Artificial Intelligence and Deep Learning in the Diagnosis of Pythium Keratitis

Etiology

Pythium insidiosum is a filamentous aquatic oomycete, an organism taxonomically belonging to the kingdom Stramenopila, class Oomycota, and more closely related to diatoms and brown algae than to true fungi. This evolutionary lineage explains its unique biochemical profile and clinical behavior. Unlike fungi, Pythium lacks chitin and ergosterol and instead uses cellulose and β-glucans in its cell wall. This biochemical difference renders conventional antifungal drugs, such as amphotericin B, natamycin, and azoles, largely ineffective because their primary targets (ergosterol or chitin synthesis) are absent.[13]

The organism’s life cycle centers around the production of biflagellate zoospores, which thrive in stagnant freshwater, paddy fields, irrigation channels, aquaculture tanks, and wet soil. These motile zoospores exhibit chemotactic attraction toward damaged corneal epithelium. Upon reaching the ocular surface, they encyst, adhere firmly to the corneal substrate, and germinate, forming invasive hyphae that penetrate the Descemet membrane and invade deeper stromal layers.[14]

Once inside the cornea, Pythium secretes a variety of hydrolytic enzymes, including cellulases, proteases, lipases, and collagenases, which collectively degrade the stromal extracellular matrix. This enzymatic degradation leads to rapid stromal necrosis, ring infiltration, and the characteristic tentacle-like radial extensions visible on slit-lamp examination. Recent molecular profiling has identified cellulose synthase genes and serine protease families as major virulence determinants contributing to corneal invasion.[15]

In contrast to most opportunistic keratitis pathogens, Pythium insidiosum infections commonly occur in immunocompetent individuals, often following minor ocular trauma, particularly with vegetative or aquatic material. Occupational exposure plays a crucial role—farmers, fishermen, and irrigation workers are at the highest risk. Environmental factors such as warm temperature (28–35 °C), high humidity, and stagnant water reservoirs enhance zoospore proliferation and persistence, explaining the seasonal spikes during monsoon months in tropical Asia.[16] Interestingly, more recent epidemiologic data also show Pythium infections among contact lens wearers and post-keratoplasty recipients, which are attributed to microabrasions, biofilm formation, and contaminated lens solutions. These cases highlight the pathogen’s adaptability beyond agrarian exposures.

Artificial intelligence (AI) is now playing an important role in etiologic characterization. Deep learning algorithms can analyze cytologic smear and histopathology images to distinguish Pythium hyphae from those of true fungi. This approach has enabled automated screening for Pythium in digital pathology, which is especially useful in laboratories with limited mycology expertise. Additionally, machine-learning models have been applied to environmental surveillance. Remote sensing data and humidity/temperature indices, integrated with AI-based predictive analytics, have been proposed for forecasting outbreaks in endemic agricultural regions. Such innovations highlight how AI can bridge basic microbiological understanding and real-world epidemiologic monitoring.[17]

Table. Etiologic and Environmental Determinants of Pythium insidiosum Keratitis

Table. Pathogenic and Diagnostic Determinants Influencing the Severity and Management of Pythium insidiosum Keratitis

Epidemiology

Pythium insidiosum keratitis (PIK) has evolved from a regional tropical infection into a global ophthalmic concern, reflecting both climatic adaptability and advances in diagnostic recognition (see Image. Schematic diagram depicting epidemiology of Pythium insidiosum keratitis). First described in Thailand and India in the late 20th century, it was long mistaken for fungal keratitis. Over the past decade, the pathogen has been increasingly reported from South and Southeast Asia, Oceania, the Middle East, and even North America, marking its emergence as a pathogen of international relevance.[18]

Global Distribution

• India: India currently represents the largest global burden of PIK, with endemic clusters reported from Tamil Nadu, Kerala, Madhya Pradesh, Gujarat, and Rajasthan. Gurnani et al (2023) noted a substantial increase in Pythium isolation rates during the monsoon season, with 12% to 15% of initially culture-negative microbial keratitis cases subsequently confirmed as Pythium by histopathology or molecular assays. Hotspot mapping indicates that western and southern India are at particular risk due to rice cultivation, exposure to stagnant water, and humid climates.[19]
• Thailand: Thailand historically contributed the earliest case series, especially among rice-field workers and irrigation farmers. Pythium remains endemic in the Chiang Mai and Bangkok regions, with community-based surveillance studies reporting 3%–5% of microbial keratitis isolates attributable to Pythium.[20]
• Australia: Sporadic outbreaks have been documented in northern and northeastern Australia, particularly in Queensland and the Northern Territory, often associated with exposure to floodwater and warm coastal climates.[21]
• United States: Although rare, Pythium keratitis has been reported in Florida, Texas, and Louisiana, typically associated with aquatic, horticultural, or domestic animal exposures. The presence of Pythium in temperate climates suggests global dissemination facilitated by waterborne spore persistence and climatic shifts.[22]
• Middle East and Africa: Isolated reports from Saudi Arabia, Egypt, and sub-Saharan Africa highlight the possibility of under-recognition in arid zones due to diagnostic limitations rather than true absence.

Demographics and Risk Factors

PIK affects both sexes almost equally, although a slight male predominance has been observed (M:F distribution is approximately 1.3:1), likely reflecting occupational exposure in agrarian settings. The median age group is 35–50 years, corresponding to the working rural population. Pediatric cases are uncommon but documented; these often follow trauma or contaminated water exposure. The infection is not associated with immunodeficiency, differentiating it from fungal and bacterial opportunistic keratitis.[23]

Seasonal peaks correlate strongly with monsoon months (June–September) in India and Thailand, paralleling increased rainfall, irrigation activity, and freshwater contamination. Temperature thresholds between 28 to 35 °C and high relative humidity (>80%) favor zoospore motility and encystment (see Image. Schematic diagram depicting risk factors for Pythium insidiosum keratitis).

Environmental Correlates and Climate Dynamics

Environmental surveillance studies have detected Pythium DNA in rice-field water, wetlands, aquaculture ponds, and soil sediments, confirming its saprophytic aquatic lifestyle. Seasonal flooding, irrigation, and stagnant water facilitate spore proliferation and ocular exposure through splashing or trauma. Climate change and rising global temperatures are likely to expand the pathogen’s ecological range, with increasing reports from subtropical and temperate latitudes. Geospatial modeling suggests that areas undergoing agricultural intensification and erratic rainfall cycles may become emerging foci.[24]

Artificial Intelligence and Predictive Epidemiology

Recent research integrates AI and geospatial machine learning to forecast Pythium outbreaks. AI-driven epidemiological mapping utilizes satellite-derived data on rainfall, soil moisture, land use, and agricultural density to model spatiotemporal disease risk. These systems employ gradient-boosted decision trees (GBDTs), convolutional neural networks (CNNs), and geographic information systems (GIS) to detect environmental signatures predictive of Pythium activity.[25]

Table. Global Epidemiology and Demographic Trends in Pythium Keratitis

Table. Geographic and Demographic Distribution of Pythium Keratitis

Pythium insidiosum keratitis exhibits a distinctive eco-epidemiological pattern, tightly linked to water, climate, and occupational exposure. Although predominantly a tropical agrarian infection, its detection in temperate countries underscores its increasingly global relevance. With rising awareness, diagnostic precision, and AI-driven surveillance, epidemiological data are being redefined from passive recognition to predictive modeling. Integrating AI-based early warning systems with clinical teleophthalmology networks promises a transformative shift toward preemptive public health measures in managing Pythium outbreaks worldwide.[15]

Pathophysiology

Pythium insidiosum keratitis (PIK) is caused by an aquatic oomycete that exhibits a distinctive pathobiologic cascade, beginning with zoospore adhesion, corneal invasion, and stromal destruction, culminating in severe inflammatory damage (see Image. Schematic diagram depicting pathophysiology of Pythium insidiosum keratitis). The organism’s virulence stems from a combination of a unique cell wall composition, enzymatic aggression, immune evasion, and biofilm formation, which together lead to rapid corneal necrosis and high post-surgical recurrence rates. Unlike fungal keratitis, the disease’s progression is driven by cellulose-mediated adhesion and collagenolytic degradation, rather than chitin-ergosterol-based mechanisms.

1. Entry and Initiation of Infection

The infection begins when biflagellate zoospores encounter a compromised corneal epithelium, typically following trauma from vegetation or aquatic exposure. These motile spores exhibit chemotaxis toward damaged corneal surfaces, guided by chemical cues, including amino acids, sugars, and ionic gradients generated by tear film proteins. Upon contact, zoospores encyst, shed their flagella, and form a glycoprotein capsule that facilitates adhesion to epithelial cells and stromal collagen.

AI-assisted microscopic analysis of early-phase infections has shown that Pythium zoospores preferentially attach to epithelial microdefects and Descemet membrane microfolds, which can now be digitally mapped through AI-enhanced confocal microscopy. Such imaging reveals distinct “gliding” patterns of encysted cells, unlike the budding seen in fungi.[26]

2. Germination and Hyphal Invasion

Once encysted, germ tubes emerge within 1 to 2 hours, forming coenocytic (aseptate) hyphae that penetrate the corneal stroma. These hyphae exhibit broad, ribbon-like morphology (4–10 µm width) and show strong tropism toward collagen-rich zones. The infection rapidly spreads laterally through stromal lamellae, aided by the following enzymatic mechanisms.

• Cellulase and β-1,3-glucanase enzymes break down stromal polysaccharides.
• Collagenase and protease complexes degrade the extracellular matrix (ECM).
• Lipases and phospholipases damage host cell membranes.

AI-based segmentation models applied to digital histopathology slides can distinguish these enzymatic zones by analyzing stromal reflectivity and necrosis gradients, correlating with clinical severity scores.[27]

3. Host Immune Response

PIK elicits an acute neutrophilic immune response, unlike the mixed granulomatous inflammation typical of fungal infections. Neutrophils release reactive oxygen species (ROS) and proteolytic enzymes in an attempt to contain the infection, but the pathogen’s cellulose capsule resists phagocytosis. The organism then secretes serine protease inhibitors and elastase analogues that neutralize neutrophil elastase and impair oxidative killing.

This immune dysregulation leads to the formation of dense stromal abscesses, peripheral ring infiltration, and endothelial plaques. In vivo confocal microscopy (IVCM) reveals hyper-reflective hyphal strands with surrounding activated keratocytes and leukocytes, forming the diagnostic “reticular honeycomb” pattern. AI-aided image analysis allows quantitative grading of inflammation by measuring reflectivity, cell density, and stromal edema, enabling objective staging.[28]

4. Enzymatic and Molecular Pathways

Molecular studies have identified multiple virulence-associated genes in Pythium insidiosum:

• PiC1 and PiC2 encode cellulase isoenzymes responsible for stromal penetration.
• PiP1 (Protease-1) and PiCol (Collagenase) promote degradation of the Bowman and Descemet membranes.
• PiA1 (Adhesin) facilitates biofilm adherence to corneal stroma.
• PiGly (Glycosyl hydrolase) enhances nutrient acquisition in hypoxic conditions.

AI-driven molecular docking and in silico modeling studies have further elucidated enzyme–substrate interactions, predicting strong binding affinities between cellulose synthase inhibitors (CBIs) (eg, carpropamid) and the PiC1 active site. These computational insights are now guiding drug repurposing efforts for non-antifungal PIK therapy.[29]

5. Stromal Necrosis and Corneal Destruction

As enzymatic activity intensifies, stromal lamellae lose tensile integrity, leading to melting and descemetocele formation. This necrotic stage often develops within 7 to 10 days from onset. AI-assisted AS-OCT imaging quantifies stromal reflectivity loss and edema index, enabling early prediction of impending perforation. The characteristic “tentacular extensions” seen clinically correspond to hyphal migration tracks, which appear as linear hyperreflective projections on AS-OCT and IVCM. CNN-based texture analysis of these tentacles achieves over 95% sensitivity in distinguishing Pythium from fungal infiltrates, making it a noninvasive biomarker.[30]

6. Resistance to Antifungal Therapy

The absence of ergosterol in Pythium membranes explains the complete ineffectiveness of amphotericin B, natamycin, and azoles. Furthermore, the organism’s cellulose-based membrane and high β-glucan content limit permeability to polyene and azole molecules. The biofilm matrix impedes drug diffusion, while enzymatic metabolism neutralizes drug activity.[31]

7. Post-Keratoplasty Recurrence Mechanisms

Recurrence after therapeutic penetrating keratoplasty (TPK) occurs in 20% to 40% of cases. The residual infection persists in the host scleral margins or limbal tissues. Histopathologic sections of failed grafts reveal dormant hyphal fragments surrounded by fibrovascular proliferation.

AI-assisted postoperative slit-lamp surveillance can automatically detect micro-recurrence patterns (eg, minute satellite infiltrates or texture irregularities) weeks before clinical recognition. This early detection allows prompt re-intervention, thereby significantly improving graft survival.[32]

8. Artificial Intelligence in Pathophysiologic Modeling

AI technologies have begun reshaping the understanding of PIK pathogenesis beyond microscopy. Three key domains stand out:

1. Image-Based Pathology Recognition CNNs, Vision Transformers (ViTs), and autoencoders trained on corneal histopathology images can now identify Pythium hyphae, quantify enzymatic necrosis zones, and predict organism viability index.
2. Predictive Disease Modeling Machine-learning algorithms integrate patient demographics, environmental exposure, and imaging biomarkers to model infection kinetics. Random forest models can predict time to perforation, need for TPK, and likelihood of recurrence with high precision.
3. Molecular Simulation and Drug Targeting AI-driven molecular dynamics and generative models simulate enzyme–drug interactions, identifying new potential inhibitors targeting cellulose synthase and serine protease pathways. These in silico platforms accelerate therapeutic discovery without the need for prolonged culture-based testing.[33]

9. Neuro-Immune Crosstalk and Pain Mechanisms

PIK also exhibits atypical pain profiles—often disproportionate to corneal findings. This finding is attributed to neuroimmune inflammation, in which proinflammatory cytokines (IL-1β, TNF-α) released by keratocytes sensitize corneal nociceptors. AI-based ocular-surface thermography and optical flow analysis can quantify inflammation-induced microvascular dilation and correlate pain intensity with disease activity.[34]

10. Systemic Correlation and Future Directions

Although primarily ocular, disseminated Pythium infection has been reported in immunocompromised hosts, particularly involving the skin or arteries. Understanding the pathophysiology of ocular disease helps identify systemic manifestations earlier.

Future research integrating AI-based 4D modeling of stromal degradation with real-time confocal video analytics could revolutionize early detection and risk prediction, offering dynamic insights into pathogen–host interactions.

In summary, the pathophysiology of Pythium insidiosum keratitis is characterized by rapid stromal destruction driven by enzymatic virulence and immune dysregulation, compounded by antifungal resistance due to its unique oomycete cell wall composition. Artificial intelligence provides unprecedented tools for visualizing, quantifying, and predicting these microscopic processes, bridging the gap between bench-level microbiology and real-time clinical decision support. By combining deep-learning image analytics, molecular modeling, and predictive algorithms, AI is transforming Pythium keratitis from a diagnostic challenge into a model of data-driven ocular pathology.[15]

Table. Comparison of Structural and Functional Pathophysiology: Pythium vs Fungal Keratitis

Histopathology

Histopathological evaluation remains the cornerstone for confirming Pythium insidiosum keratitis (PIK), particularly when culture or PCR results are inconclusive (see Image. Schematic diagram depicting histopathology of Pythium insidiosum keratitis). The microscopic features of Pythium infection closely mimic those of filamentous fungal keratitis, but subtle morphological and staining differences—especially when interpreted through digital or AI-assisted microscopy—enable accurate distinction.

Gross and Microscopic Morphology

The corneal tissue involved in PIK shows dense stromal infiltration, necrosis, and loss of normal lamellar architecture. The hallmark finding is the presence of broad, ribbon-like, sparsely septate hyphae (width: 3–10 µm) that infiltrate throughout the corneal stroma and sometimes extend up to the Descemet membrane or anterior chamber angle (see Image. Schematic diagram depicting histopathology of Pythium insidiosum keratitis). These hyphae are poorly refractile, with irregular branching at right or obtuse angles—unlike the regular dichotomous branching seen in Aspergillus or Fusarium infections. The organism typically demonstrates a paucity of septa, though occasional pseudo-septa may appear due to degenerative changes.[2]

The inflammatory response is predominantly neutrophilic, with scattered necrotic keratocytes and fibrin deposition. Epithelial ulceration, stromal edema, and microabscess formation are frequent. The absence of granulomatous inflammation, a common feature in fungal infections, is diagnostically relevant. In chronic stages, fibrovascular proliferation and stromal scarring may be seen around degenerated hyphal remnants.

Staining Characteristics

Routine hematoxylin and eosin (H&E) staining shows Pythium filaments as pale, eosinophilic, poorly staining hyphae, often surrounded by neutrophilic debris.

• Periodic acid–Schiff (PAS): Weak or variable staining due to cellulose-based walls lacking chitin.
• Gomori methenamine silver (GMS): Stains filaments faintly gray to black, but less intensely than true fungi.
• Calcofluor white: Fluorescent staining highlights the cellulose-rich walls, producing a linear “ribbon” pattern under UV microscopy.
• Gram stain: Variable positive reaction; Gram-negative in degenerative phases.

Digital histopathology employing AI-based segmentation can now quantify staining intensity and filament morphology, reducing observer bias. Automated algorithms using convolutional neural networks (CNNs) classify hyphal structures and stain uptake, achieving over 93% accuracy in differentiating Pythium from fungal elements.[15]

Tissue Localization and Spread

Hyphal elements are usually concentrated in the mid to deep stroma, often clustering along collagen lamellae. In advanced cases, they penetrate the Descemet membrane, invade the endothelium, and may enter the anterior chamber, forming exudative plaques or hypopyon. Perineural invasion is rare, distinguishing PIK from Acanthamoeba keratitis. Endothelial involvement correlates with poor prognosis and early need for keratoplasty.

AI-driven image analysis of whole-slide scans can measure hyphal density, invasion depth, and inflammatory zone thickness, providing objective metrics for prognostic stratification. Deep-learning models trained on annotated histopathology slides can automatically map these features and calculate a “pathogen burden index” that correlates strongly with clinical severity.[35]

Immunohistochemistry (IHC) and Molecular Correlation

IHC enhances diagnostic specificity when traditional stains are inconclusive. Antibodies raised against Pythium insidiosum cell-wall components—particularly cellulose synthase (PiCS) and β-glucan epitopes—show strong cytoplasmic reactivity.

Lectin-based assays (eg, concanavalin A binding) differentiate Pythium from fungal hyphae by identifying mannose-deficient cell wall components.

Complementary PCR amplification of internal transcribed spacer (ITS) and cytochrome oxidase II (COX2) gene fragments from histologic sections confirms identity, especially when culture fails. Histology–molecular correlation remains the gold standard for diagnosis, with a reported concordance of 90% to 95%.

AI-assisted IHC interpretation, through pixel-level pattern recognition, reduces interobserver variability in immunoreactive area quantification. Emerging models also use spectral imaging AI to detect Pythium-specific chromogenic signatures that are invisible to the human eye, thereby enabling earlier histopathologic alerts within digital workflows.[36]

Comparison with Fungal Keratitis

The distinction between Pythium and true fungi is subtle but critical, as it directly dictates therapeutic choice. AI-based differential diagnosis tools using digital histopathology (ResNet50, EfficientNet) can distinguish these categories by extracting morphometric features, reducing misclassification and improving early therapeutic decision-making.

Table. Histopathologic Differentiation Between Pythium and Fungal Keratitis*

*AI performance metrics vary by dataset and architecture; convolutional neural network (CNN)–based models have reported classification accuracies of approximately 92%–95% in differentiating Pythium from fungal keratitis in pilot studies.

Role of Confocal and Digital Pathology

In vivo confocal microscopy (IVCM) complements histopathology by providing real-time, noninvasive visualization of Pythium filaments. Characteristic IVCM findings include:

• Linear, hyperreflective strands with right-angled branches
• Reticular “net-like” stromal pattern
• Peripheral radial extensions mimicking “tentacles”

Deep-learning–assisted IVCM interpretation using CNN-based models enables automatic segmentation and hyphal density scoring, which correlate with histologic findings. Combined AI pipelines linking IVCM and histopathology datasets can thus create comprehensive diagnostic frameworks, shortening time-to-diagnosis from days to hours.[37]

Post-Keratoplasty Histopathologic Changes

In post-therapeutic keratoplasty specimens, Pythium hyphae are often localized at graft-host junctions, occasionally extending into scleral tissue. Fibrovascular proliferation, residual inflammatory infiltrate, and stromal remodeling are typical. AI-aided analysis of these sections helps differentiate active infection (vital hyphae with cytoplasmic granularity) from post-treatment scarring (degenerate or ghost hyphae), crucial for postoperative management and antimicrobial withdrawal decisions.[38]

AI and Digital Histopathology Workflow

The integration of AI into corneal histopathology has revolutionized the efficiency of interpretation.

• Segmentation algorithms automatically delineate corneal layers and quantify hyphal load.
• Classification models distinguish Pythium, fungal, and bacterial etiologies based on texture and morphology.
• Heatmap visualization highlights regions of diagnostic significance for pathologists.
• Cloud-based validation enables remote digital pathology review via teleophthalmology platforms.[39]

In summary, histopathology in Pythium insidiosum keratitis provides the most definitive morphological evidence of infection. The presence of broad, sparsely septate, cellulose-rich filaments, weak special staining, and neutrophilic infiltration differentiates it from fungal keratitis. With the advent of AI-assisted digital pathology, diagnostic precision has improved dramatically, allowing automated hyphae recognition, pattern quantification, and outcome prediction. The convergence of classical microscopy and computational pathology ensures that Pythium keratitis—once easily misdiagnosed—can now be identified with near–real-time accuracy, facilitating timely therapeutic interventions and improving visual outcomes.[2]

Toxicokinetics

Although “toxicokinetics” traditionally pertains to drug metabolism, its application in PIK underscores why antifungals fail and how AI models may predict therapeutic response. Because Pythium lacks ergosterol, polyenes (eg, amphotericin B) cannot bind effectively, and azoles (eg, voriconazole, itraconazole) fail to inhibit 14-α-demethylase. Computational pharmacokinetic modeling has revealed poor drug–organism affinity scores.

AI-assisted in-silico docking and pharmacovigilance systems can simulate molecular interactions, predict potential inhibitors (such as cellulose biosynthesis inhibitors [CBIs]), and estimate corneal penetration based on physicochemical parameters. Deep-learning regression models trained on ocular drug datasets can forecast intra-stromal drug concentration and toxicity risks, thereby guiding drug repurposing.[40]

Table. Comparative Drug Sensitivity and AI-Predicted Efficacy

History and Physical

Patients with Pythium insidiosum keratitis (PIK) typically present with acute or subacute onset of pain, redness, watering, and progressive visual loss following exposure to contaminated or stagnant water, especially in agricultural or monsoon settings (see Image. Schematic diagram depicting clinical features of Pythium insidiosum keratitis). The disease is often misdiagnosed as fungal keratitis, leading to delayed or inappropriate antifungal therapy.

Key Historical Features

• History of exposure: Bathing, swimming, or working in rice fields, ponds, or muddy water
• Preceding trauma: Minor vegetative or soil-related corneal injury in approximately 70%–80% of cases
• Treatment delay: Often self-treated or treated with antifungals, worsening the prognosis
• Demographics:
• Most common in young to middle-aged males involved in agriculture
• Pediatric and elderly cases are less frequent but possible
• Duration:
• Symptoms develop within 3 to 5 days post-exposure and progress rapidly despite antifungal therapy.[41]

Common symptoms:

• Severe ocular pain disproportionate to lesion size
• Photophobia, foreign body sensation, and profuse lacrimation
• Diminished or blurred vision due to dense stromal infiltration
• Non-healing ulcer despite antifungal use—an immediate red flag for Pythium[42]

Physical Examination (Ocular Findings)

General appearance:

• Gray-white, dry-looking, reticular stromal infiltrate with ill-defined, feathery or tentacular extensions—a hallmark of Pythium keratitis
• Lesions typically located in the paracentral or peripheral cornea with surrounding guttering and minimal stromal edema[43]

Disease course on examination:

• Early stage: Small anterior stromal infiltrate with intact endothelium and no hypopyon
• Progressive stage: Mid-stromal spread with reticular/tentacular extensions and early hypopyon
• Advanced stage: Deep stromal melt, endothelial plaque, and limbal involvement; risk of perforation is high[44]

Characteristic clinical clues suggesting PIK:

1. Non-resolving “fungal-looking” keratitis unresponsive to antifungals within 3 to 4 days
2. Reticular infiltrate with tentacle-like radial extensions
3. Dry, rough, elevated stromal lesion with minimal surrounding edema
4. Absence of pigmentation (unlike dematiaceous fungi)
5. Rapidly progressive ulcer despite antifungal therapy adherence[17]

Systemic findings:

• PIK is usually limited to ocular involvement, but severe ocular inflammation can cause secondary systemic symptoms such as headache or periocular pain

• Systemic dissemination is not seen in immunocompetent individuals.

Table. Characteristic Clinical Findings in Pythium insidiosum Keratitis

Evaluation

Evaluation of Pythium insidiosum keratitis (PIK) requires an integrated, multimodal approach that combines clinical imaging, microbiology, molecular assays, and AI-enabled analytics. PIK should be suspected in any rapidly worsening, antifungal-resistant keratitis in a patient with agricultural or aquatic exposure. Clinical hallmarks include a gray reticular infiltrate with tentacular margins, severe pain, endothelial plaque, and early limbal spread. Timely recognition at the bedside, reinforced by AI-assisted imaging and microbiologic confirmation, remains the cornerstone of sight-saving management.[45]

Conventional Methods

• Corneal scraping and smear: Detection of broad aseptate filaments on KOH, Gram, and calcofluor staining
• Culture: Slow growth on chocolate or blood agar within 5 to 7 days
• PCR/LAMP: Highly specific detection of Pythium DNA targets (ITS and COX2 genes)
• IVCM/AS-OCT: Visualization of hyperreflective stromal strands and tentacular extensions [46]

AI-Based Diagnostics

• Deep learning now provides rapid, reproducible classification directly from slit-lamp or confocal images.
• ResNet-50, DenseNet-121, and EfficientNet models have achieved AUC >0.95 in multiclass studies.
• Algorithmic workflow: image capture → pre-processing (normalization, ROI segmentation) → feature extraction → classification → output probability map [47]

Molecular and biosensor integration

Recent prototypes merge biosensors that detect Pythium antigens or cellulose derivatives with AI-based signal interpretation, enabling on-site point-of-care testing within 30 minutes.

Validation and regulatory guidelines

The Indian Council of Medical Research (ICMR) and the Thai Ophthalmic Society have endorsed multimodal evaluation strategies that emphasize AI-assisted triage for culture-negative keratitis. International AI guidelines (AAO AI Task Force, 2023) recommend explainable models and data transparency.[48]

Clinical Screening and Image Acquisition (input for AI and baseline care)

• Triage triggers suggesting PIK
  • Rapidly progressive stromal infiltrate with dry, reticular, or "tentacular" margins
  • Peripheral guttering/advancing edge
  • Minimal endothelial plaque
  • Feathery hyphal-like streaks but poor response to antifungals
  • Exposure to paddy/wet soil/floodwater 
  • Monsoon season
• Standardized slit-lamp imaging protocol (vital for AI)
• Diffuse, focal, sclerotic scatter, and retro-illumination views; include scale bar and white balance
• At least 3 projections (central, nasal, temporal) and 1 anterior segment video, if possible
• Record ulcer size (H × V in mm), depth, hypopyon height, and peripheral extension
• For tele-triage or field work: smartphone + slit-lamp adaptor (≥12 MP), fixed 1:1 macro, no digital zoom
• In vivo confocal microscopy (IVCM), where available:
  • Long, thin, branching aseptate filaments (3–7 μm)
  • Right-angle branching patterns
  • Segmental hyperreflective swellings (“string-of-beads” appearance)
  • Deeper stromal sheet-like infiltrates
  • Technical note: Save raw image stacks (TIFF format) for AI-based analysis [6]

Microbiology and Pathology

(Ground truth for AI labels and routine diagnosis)

Essential smears

(Bedside, within 30 min of scraping)

• 10% potassium oxide (KOH) ± calcofluor white: Broad, sparsely septate filaments
• Gram stain: Weakly gram-positive hyphae (useful for excluding bacterial infection)
• Iodine–potassium iodide plus 1% sulfuric acid (IKI–H2SO4): Highlights the cellulose-rich cell wall of Pythium [49]
• Giemsa stain: Adjunctive staining

Culture

(Minimum of two inoculation sites)

• Primary media: Blood agar (preferred), chocolate agar
• Supplementary media: Sabouraud dextrose agar (often negative); non-nutrient agar with grass leaf incubation for zoospore induction (diagnostic)
• Colony morphology: Flat, colorless colonies with a radiating growth pattern
• Microscopy: Identification of sporangia and motile zoospores

Molecular and rapid diagnostic tests

(Where available)

• PCR: Targeting ITS rDNA, with sequencing for species confirmation
• qPCR or LAMP: Rapid detection suitable for low-resource laboratory settings
• MALDI-TOF MS: Identification from culture isolates, where locally validated[50]

Histopathology

(Therapeutic penetrating keratoplasty or evisceration specimens)

• Hematoxylin and eosin (H&E): Broad, sparsely septate filaments within the stroma
• Gomori methenamine silver (GMS): Negative or weak staining
• Periodic acid–Schiff (PAS): Variable positivity
• Calcofluor white: Strong positivity
• Cell wall composition: Cellulose-rich walls
• Angioinvasion: Uncommon

These laboratory findings constitute the reference standard for training and validation of AI-based diagnostic systems.

Artificial Intelligence and Deep Learning Evaluation Pipeline

Data curation & labeling

• Data sources: Slit-lamp photos ± IVCM image stacks from microbiologically confirmed Pythium, fungal, bacterial, Acanthamoeba, and herpetic keratitis cases
• Label derivation: Reference labels assigned based on culture, PCR, and/or histopathology (not treatment response), with annotation of onset-to-image time and pretreatment status
• Dataset partitioning: Patient-level separation with inclusion of an external-site test set (eg, different hospital, geographic region, or season)
• Quality control: Exclusion of overexposed or blurred images, with retention of a distinct “ungradable” category[51]

Model architectures

• Classification models: ResNet, EfficientNet, Vision Transformer architectures for binary to multiclass classification tasks (eg, PIK vs fungal vs bacterial vs other keratitis)
• Segmentation models: U-Net or Mask R-CNN architectures for delineation of stromal infiltrates and tentacular extensions, enabling surrogate estimates of lesion area and depth
• Multimodal fusion: Integration of image-derived features with clinical metadata (eg, trauma history, water exposure, symptom duration) and IVCM features to enhance diagnostic sensitivity

Performance reporting

(Aligned with TRIPOD-AI and CONSORT-AI principles)

• Primary metrics: Sensitivity for PIK detection, area under the receiver operating characteristic curve (AUC), and F1 score
• Secondary metrics: Specificity against filamentous fungal keratitis, calibration (Brier score), and decision-curve analysis (net benefit at clinically relevant TPK referral thresholds, eg, ≥0.7 predicted probability)
• Robustness analyses: Subgroup performance by camera type, clinical center, ulcer size, and pretreatment status[52]

Human–AI clinical workflow

1. Image acquisition: Capture standardized slit-lamp images and process them through on-device or server-based AI models
2. Triage output interpretation:
• High probability PIK (≥0.7): Urgent microbiologic evaluation and IVCM; initiation of anti-oomycete therapy (avoiding azoles); early TPK consultation
• Intermediate probability (0.40–0.69): Repeat imaging; expanded laboratory testing (eg, grass leaf incubation); consideration of qPCR
• Low probability (<0.40): Management according to standard microbial keratitis pathways, without deferral of smear or culture
3. Explainability: Visualization using gradient-weighted class activation mapping (Grad-CAM) or heatmaps highlighting reticular margins and tentacular extensions, archived with the diagnostic report[53]

Deployment and safety checks

• Edge deployment: Progressive web applications or Android-based platforms supporting offline inference and automated deidentification
• Model governance: Locked model versions, audit logging, and scheduled recalibration to account for seasonal or epidemiologic drift
• Human oversight: Mandatory clinician review, with AI outputs functioning as adjunctive decision support rather than replacements for microbiologic diagnosis

Minimum diagnostic dataset at first clinical encounter

• Slit-lamp photography (≥3 standardized views), ulcer dimensions, and hypopyon height
• Corneal scrapings for KOH with calcofluor white, Gram stain, and IKI–H2SO4
• Culture on blood agar with adjunctive grass leaf incubation
• IVCM, where available
• qPCR or LAMP, where accessible
• AI-based image triage (if available), with probability estimates and corresponding heatmaps[52]

Criterion for escalation and early therapeutic keratoplasty (TPK)

• AI-predicted PIK probability ≥0.7 or classic morphologic features with poor or absent response to antifungal therapy within 48–72 hours
• Rapid peripheral progression, involvement of more than two limbal quadrants, deep stromal sheet-like infiltrates on IVCM, or impending/perforated cornea
• Coordination with microbiology services for culture of the excised corneal button (critical for reference labeling and future AI training)

Documentation & reporting standards for AI-enabled clinics

• Inclusion of AI-derived probability scores, model version identifiers, and heatmaps in the clinical record
• Documentation of time from presentation to definitive laboratory confirmation as a quality assurance metric
• Flagging and periodic review of AI–laboratory discordance to inform active learning datasets[54]

Quality, safety, and regulatory considerations

• Established clinical guidelines for microbial keratitis remain foundational (institutional, AIOS, or RCOphth pathways), including prompt smear and culture acquisition, avoidance of exclusive antifungal therapy in suspected PIK, and early consideration of TPK when progression occurs despite optimized treatment
• AI development and deployment should adhere to recognized frameworks, including:
• TRIPOD-AI for model reporting and CONSORT-AI/SPIRIT-AI for clinical trials
• Software as a medical device (SaMD) principles from the IMDRF and FDA; alignment with CDSCO digital health guidance and institutional review board approval in India
• WHO's ethical guidance for AI, including transparency, data minimization, bias monitoring, and human oversight
• Data governance: Explicit imaging consent, rigorous deidentification, site-level data-sharing agreements, and periodic bias audits (eg, rural vs urban settings, device types, seasonal variation)[55]

Suggested diagnostic evaluation algorithm (textual)

Suspect PIK → standardized imaging → smear panel (KOH, Gram, IKI-H2SO4) and cultures (blood agar ± grass leaf incubation) ± IVCM → AI-based image analysis

• High-probability or classic morphology: Anti-oomycete pathway, early TPK consultation, qPCR/LAMP
• Indeterminate: → repeat imaging, broaden labs, consider referral.
• Low probability with persistent clinical concern: Management per microbial keratitis protocols without delaying laboratory evaluation; reassessment at 24–48 hours

Recommended reporting elements

• Imaging dataset and views
• AI probability scores and heatmaps
• Smear, culture, and molecular results
• IVCM descriptors
• Clinical decision-making rationale, including TPK referral
• Follow-up plan [56]

Table. AI and Imaging Correlates (Modern Diagnostic Support)

Table. Distinguishing Features of Pythium insidiosum Keratitis and Fungal Keratitis

Treatment / Management

Pythium insidiosum keratitis (PIK) requires an urgent, multidisciplinary approach that differs fundamentally from that for fungal keratitis. Because Pythium is an oomycete and not a fungus, conventional antifungal drugs are ineffective. Optimal management integrates early anti-oomycete antibiotic therapy, timely surgical intervention, and close postoperative surveillance guided by both clinical staging (Gurnani–Kaur) and AI-assisted diagnostic inputs.[57]

Principles of Management

• Avoid antifungal monotherapy:  Pythium lacks ergosterol and chitin, rendering antifungals (eg, natamycin, voriconazole, amphotericin B) ineffective.
• Initiate anti-oomycete therapy early: High-frequency combination therapy with linezolid and azithromycin is recommended.
• Avoid corticosteroids during the active phase: Corticosteroids may accelerate stromal destruction and increase the risk of recurrence.
• Reassess clinical response at 48–72 hours: Progression or lack of improvement warrants early therapeutic penetrating keratoplasty (TPK).
• Ensure wide surgical excision margins: Margins of ≥1–1.5 mm reduce the risk of recurrence.
• Incorporate AI-based image tracking when available: Serial analysis may assist in monitoring treatment response and predicting progression.[58]

Medical Management

Table. First-line Topical Regimen (Anti-Oomycete Protocol) for Pythium Keratitis

Table. Systemic Therapy (Adjunctive) for Pythium Keratitis

Response Evaluation (48–72 hours):

• Reduction in stromal infiltrate size or density
• Absence of new tentacular extension
• Improvement in pain, conjunctival hyperemia, and hypopyon
• Lack of improvement or evidence of progression warrants escalation to surgical management.[59]
(B3)

Surgical Management

Indications for therapeutic penetrating keratoplasty (TPK)

• Infiltrate >6 mm, >2 quadrants, or approaching limbus
• Deep stromal or endothelial involvement (on slit-lamp/IVCM)
• Hypopyon >1 mm or descemetocele formation
• No clinical improvement within 48–72 hours of optimized therapy
• Impending or actual perforation[60]
(B3)

Table. Surgical Principles for the Management of Pythium Keratitis

Post-TPK follow-up

• Daily slit-lamp examination for 1 week, then every 2–3 days for one month.
• Assess for recurrence at the graft–host junction, which typically appears as a gray line or infiltrate within 2 to 3 weeks.
• In the event of recurrence, repeat TPK with a larger excision and extended margins.
• Introduce low-potency topical steroids only after confirmed infection control.[38]

Table. Role of Artificial Intelligence (AI) and Deep Learning in Management of Pythium Keratitis

Table. Gurnani-Kaur Stage-Linked Management Protocol in Pythium Keratitis

National and International Guideline Perspectives

Indian and Asian experience

• First-line therapy: Combination therapy with linezolid and azithromycin is recommended.
• Surgical timing: Early therapeutic penetrating keratoplasty (TPK) is advised when lesion size exceeds 6 mm or involves more than two limbal quadrants.
• Antifungal therapy: Antifungal-only regimens should be avoided.
• Diagnostic confirmation: Grass leaf incubation and PCR are recommended for definitive diagnosis.
• Regional validation: Case series from Thailand and Australia support similar management strategies, with improved outcomes associated with early intervention.

Global perspectives

• Guideline status: No formal World Health Organization guideline currently exists; however, regional evidence supports antibacterial-based therapy combined with wide-margin surgical excision as the prevailing standard of care.
• Emerging tools: Integration of AI for rapid triage and recurrence monitoring remains under active evaluation.[61]

 Table. Prognostic Indicators During Treatment for Pythium Keratitis

Differential Diagnosis

Pythium insidiosum keratitis (PIK) presents as a rapidly progressive, necrotizing corneal ulcer often indistinguishable from fungal keratitis on slit-lamp examination (see Image. Schematic diagram depicting the differential diagnosis of Pythium insidiosum keratitis). Accurate differentiation is essential, as conventional antifungal therapy is ineffective against Pythium insidiosum. Artificial intelligence (AI) and deep learning (DL) models can enhance diagnostic precision by distinguishing PIK from morphologically similar conditions through multimodal image analysis (eg, slit-lamp, IVCM, confocal, or OCT). The sections that follow summarize the key clinical entities that most closely mimic PIK and highlight distinguishing features critical to accurate diagnosis.

1. Fungal Keratitis (Filamentous Mycoses)

Pathogens: Fusarium, Aspergillus, Curvularia, Alternaria, etc Similarity: Feathery-edged stromal infiltrate, satellite lesions, and hypopyon resembling Pythium Distinguishing features:

• Fungal hyphae: Regularly septate, branching at acute angles (45°)
• Pythium hyphae: Sparsely septate or aseptate, broad (3–7 µm), right-angle branching
• Gomori methenamine silver (GMS): Positive staining in fungi; weak or negative GMS in Pythium
• Artificial intelligence (AI) utility: Deep CNN classifiers trained on slit-lamp images show >90% AUC in distinguishing Pythium vs Fusarium; heatmap localization (Grad-CAM) highlights broader, reticular “tentacle-like” edges in Pythium.[9]

2. Bacterial Keratitis

Pathogens: Pseudomonas aeruginosa, Staphylococcus aureus, Streptococcus pneumoniae Similarity: Dense stromal infiltrate with epithelial defect and hypopyon Distinguishing features:

• Rapid onset (24–48 h), mucopurulent discharge, marked pain, and systemic response
• Pythium ulcers are usually slower (days), have a dry surface, and lack a mucopurulent reaction.
• Confocal: Bacteria are invisible; Pythium shows linear, branching filaments.
• AI utility: CNNs can achieve >95% accuracy in distinguishing between bacterial and Pythium using texture features (surface irregularity, dryness, and stromal reflectivity gradients).[62]

3. Acanthamoeba Keratitis

Similarity: Ring infiltrate, severe pain disproportionate to clinical signs, and confusion with early Pythium stages Distinguishing features:

• History of contact lens wear or water exposure
• Pain disproportionate to ulcer size
• IVCM: Double-walled cysts (highly reflective round structures) vs Pythium's linear branching filaments
• AI utility: Hybrid CNN + SVM models trained on IVCM images can distinguish Acanthamoeba cysts (spherical, highly reflective spots) from Pythium filaments (linear, non-septate strands).[63]

4. Herpetic (HSV) Keratitis

Similarity: Stromal infiltrates with ulceration and scarring may mimic late-stage PIK. Distinguishing features:

• Dendritic/geographic ulcers, decreased corneal sensation, and vesicular rash history
• Pythium lacks dendrites and has an irregular “reticular” infiltrate.
• IVCM: Inflammatory cells, but no filamentous elements in HSV
• AI utility: DL-based image segmentation can detect branching dendritic ulcers typical of HSV, reducing false positives for Pythium.[32]

5. Nocardia Keratitis

Similarity: Superficial, patchy infiltrate with satellite lesions. Distinguishing features:

• Nocardia: Characteristic “cracked glass” appearance with a raised surface; filamentous but thinner (1 µm), and weakly acid-fast on Ziehl–Neelsen
• Pythium: Poor acid-fast staining; grows readily on blood agar without branching aerial hyphae
• AI utility: Morphometric CNN features can detect Nocardia’s punctate surface texture vs Pythium’s tentacular spread.[64]

6. Mixed Microbial Keratitis

Combination: Fungal or bacterial co-infection with Pythium may occur, especially in tropical climates. Indicators:

• Mixed culture growth; inconsistent smear results
• Progressive ulcer despite combined therapy
• AI utility: Ensemble models that integrate clinical and microbiological metadata improve classification performance for mixed infections through pattern fusion.[4]

7. Noninfectious Keratitis (Sterile, Mooren’s, or Autoimmune Ulcers)

Similarity: Peripheral infiltrates and stromal melt. Distinguishing features:

• Minimal pain, non-progressive, immune origin
• Absence of microbial filaments, negative cultures
• AI utility: AI systems trained exclusively on infectious causes may misclassify sterile ulcers; algorithms should include “noninfectious” as a control class.[6]

8. Corneal Foreign Body/Trauma-Induced Necrosis

Similarity: Focal infiltrate with inflammation and necrosis. Distinguishing features:

• Clear mechanical entry point with visible metallic/vegetative debris
• Limited progression once the foreign body is removed
• AI utility: Object detection networks can localize metallic reflectivity or foreign-body shadows, thereby differentiating them from Pythium infiltrates.[65]

9. Fungal-like Oomycetes (eg, Lagenidium species)

Similarity: Morphologically similar filamentous pathogen causing keratitis or systemic infection Distinguishing features:

• Lagenidium is less commonly associated with the cornea, as confirmed by molecular sequencing (ITS)
• AI utility: AI-assisted decision trees using combined culture and image metadata can reduce misclassification.[66]

Table. Differential Diagnosis for Pythium Keratitis

In summary, Pythium keratitis can mimic bacterial and fungal ulcers both clinically and morphologically. Integrating AI-based imaging analysis, molecular confirmation, and standardized confocal or slit-lamp imaging significantly improves the accuracy of differentiation. Proper differential diagnosis prevents the use of ineffective antifungal therapy and ensures early initiation of anti-oomycete agents or therapeutic keratoplasty (TPK), thereby improving visual prognosis.[67]

Pertinent Studies and Ongoing Trials

High-quality randomized trials specific to Pythium keratitis are still limited. Most existing evidence comes from multiclass microbial-keratitis (MK) datasets in which Pythium is a prespecified class or a key subgroup. The following section synthesizes current evidence and ongoing investigations to contextualize the role of artificial intelligence as an adjunct to, rather than a replacement for, standard microbiologic diagnostics.

Table. Summary of Current Evidence Supporting the Use of AI in Pythium Keratitis

Justification for Integration of AI with Clinical Safeguards

1. Pythium is a time-critical disease that spreads peripherally; outcomes depend on early recognition and early TPK when indicated.
2. While distinct Pythium morphologies exist, they are subtle and often missed early; AI improves sensitivity during triage.
3. Net-benefit analyses show positive clinical utility at probability thresholds (eg, ≥0.7 for “treat/TPK-consult now”).
4. Explainability (Grad-CAM) helps clinicians trust the outputs by focusing on tentacular margins and reticular sheets rather than artifacts.[68]

Ongoing and Planned Clinical Trials and Registries

• Multicenter, prospective trials are comparing AI-assisted triage with standard care in suspected MK during the monsoon and harvest seasons.
  • Primary endpoints: Time from presentation to correct organism class identification (oomycete vs non-oomycete), time-to-TPK, and inappropriate antifungal days
  • Secondary endpoints: Globe salvage, best-corrected VA at 3 months, repeat-TPK rates, perforation, and cost-effectiveness
• IVCM-plus studies: Prospective accuracy studies of confocal-AI for oomycete vs filamentous fungi, with masked expert adjudication
• Tele-ophthalmology deployments (district hospitals/rural camps): On-device/PWA models for first-line triage; outcomes of interest include referral yield for probable Pythium, diagnostic turnaround time, and feasibility on mixed smartphones
• Federated-learning consortia (India/SE Asia/Australia): Site-held training to improve generalization without data leaving hospitals; outcomes are site-wise AUC and fairness across cameras and seasons
• Health-economics evaluations: Decision-analytic models using prospective service data to quantify cost per TPK averted, cost per globe saved, and opportunity costs of delayed diagnosis [69]

Recommendations for Comprehensive Evidence Reporting in AI-Enabled Programs

• Model and version; image protocol adherence (3–4 views, no digital zoom)
• Operating point used (eg, 0.70 probability = “High-probability Pythium”)
• Prospective metrics: sensitivity/specificity for Pythium, calibration (Brier), time-to-correct therapy, and net benefit
• Safety checks: proportion overridden by clinician; discrepancy audits; quarterly drift review (season/camera)
• Equity: performance by capture device, center type, and sex/age [1]

Practical Recommendations for Clinical Integration

1. Implement AI evaluation at first contact (slit-lamp photos) with a conservative, high-sensitivity threshold to trigger urgent smears, cultures, IVCM, and early TPK consult.
2. Always obtain relevant laboratory studies, including KOH + CFW, Gram, IKI-H2SO4, and culture with leaf incubation, in addition to qPCR/LAMP if available; use AI for triage and decision support.
3. Document AI probability and heatmap with final microbiology to continuously build local evidence and comulticenter multicenter registries.[66]

Treatment Planning

1) Immediate Steps at Diagnosis or Strong Clinical Suspicion

• Do not start azoles/amphotericin as sole therapy. Oomycetes respond poorly to classic antifungals.
• Run full microbiology (KOH/CFW, Gram, IKI-H2SO4), cultures (blood agar ± leaf-incubation), ± qPCR/LAMP; add IVCM if available.
• If AI or clinical morphology suggests PIK, escalate early to an anti-oomycete protocol and alert the corneal surgeon for potential early TPK.[70]

2) Medical Therapy

First-line where globe is salvageable, and there are no indications of imminent perforation

Topical antibacterials (oomycete-active):

• Linezolid 0.2% every 1 to 2 hours while awake, then taper by half each week as clinically improving
• Azithromycin 1% every 1 to 2 hours while awake, then taper similarly
• Consider oral azithromycin 500 mg once daily for 3 to 5 days (or per local protocol) and/or oral linezolid for severe or deep stromal disease. Evidence favors combination therapy over monotherapy.[71] 

Adjuncts:

• Cycloplegic (eg, atropine 1% bid) for pain/uveitis
• Anticollagenase support (eg, doxycycline 100 mg bid) as per clinician preference
• Avoid topical steroids until unequivocal control; if used post-TPK, introduce very cautiously with close monitoring.

Response assessment (24–72 h):

• Reduction in pain, epithelial defect, infiltrate density/edge activity, hypopyon; serial photos.
• Failure criteria (any): advancing tentacular edge, limbal approach, descemetocele/imminent perforation → proceed to surgical plan[72]

3) Indications for Early Therapeutic Penetrating Keratoplasty (TPK)

Early surgery improves anatomic outcomes and can prevent scleral spread. 

• Rapid peripheral spread
• Involvement of more than 2 quadrants or limbal extension
• Deep stromal sheets on IVCM
• Non-response within 48 to 72 hours
• Impending/perforated cornea.

4) Surgical Planning (TPK)

Preoperative

• Continue linezolid/azithromycin up to surgery.
• Obtain informed consent regarding the risk of recurrence and the possibility of repeat TPK.
• Plan a large, full-thickness graft to encompass all active margins.

Intraoperative principles

• Trephination margin: Plan host trephination 1–1.5 mm beyond visible disease, extending further if margins appear gelatinous or actively progressing.
• Donor graft sizing: Use a donor button 0.5–1.0 mm larger than the host bed to ensure adequate coverage.
• Host bed preparation: Irrigate the host bed and meticulously excise all infiltrated tissue. Send the excised corneal button for smear, culture, and histopathologic examination to establish definitive laboratory confirmation and provide reference-standard labels for future validation of AI-assisted diagnostic systems.
• Anterior chamber management: Remove fibrin and inflammatory debris as needed, use viscoelastic to maintain chamber stability, and secure the graft with interrupted or combined suturing to minimize wound leak.
• Assessment of scleral involvement: Evaluate for extension beyond the cornea and extend debridement as appropriate.
• Adjunctive medications: Intracameral or intrastromal drug therapy is not well established for PIK; prioritize complete excision of infected tissue.[73]

Evidence & outcomes

• Modern series show good anatomic salvage with early TPK; recurrence is possible and may require repeat, larger grafts.

Postoperative regimen

• Topical linezolid 0.2% + azithromycin 1% every 2 to 3 hours for the first week, then taper to 4 times daily over 4 to 6 weeks as it improves.
• Judicious steroid use may be introduced only after clear evidence of quiescence (no tentacular edge and negative surface cultures), balancing graft survival against recurrence risk.
• Close postoperative monitoring: Daily follow-up initially, with a low threshold for repeat scraping if recurrent infiltrate develops at the graft–host junction, as marginal recurrence is common and may require repeat TPK.[2]

5) Integration of AI-Assisted Workflow (where available)

• Use AI probability (eg, ≥0.7 for “high-prob PIK”) to trigger urgent lab confirmation and early surgeon involvement—shortens time to correct therapy and avoids futile antifungal exposure.
• Store AI heatmaps + probability in the chart; reconcile with final microbiology for QA and model calibration.

6) Disposition & Follow-Up

• Daily monitoring until clear improvement, then every 2 to 3 days.
• Educate patients on warning signs: increased pain, expansion of the gray reticular edge, new hypopyon, or wound leak.
• After epithelialization, proceed with suture management and optical rehab as appropriate; consider lamellar/optical keratoplasty later if stable.

7) Special Situations

• Mixed infection: If bacteria/fungi are also isolated, treat concomitantly while maintaining Pythium-active coverage.
• Resource-limited settings: If compounding 0.2% linezolid/1% azithromycin is difficult, liaise with regional centers; prioritize early referral for TPK.

8) Key Evidence

• Antibacterial therapy: Recent reviews recommend linezolid combined with azithromycin as first-line treatment, with the combination demonstrating greater efficacy than monotherapy.
• Adjunctive systemic therapy: Case series incorporating topical therapy with oral linezolid report cure rates exceeding 80% in selected cases.
• Role of early surgery: Early TPK is associated with high rates of anatomic globe salvage; recurrence may necessitate larger repeat grafts.
• Evolution in management strategy: Contemporary reviews emphasize a conceptual shift toward antibacterial-based regimens combined with wide-margin surgical excision.[74]

Toxicity and Adverse Effect Management

Although artificial intelligence (AI) and deep learning (DL) are diagnostic adjuncts rather than pharmacologic or radiologic interventions, “toxicity” in this context refers to algorithmic, operational, and clinical risks associated with AI use. Adverse effects may arise from model errors, data bias, over-reliance, or workflow misintegration, potentially leading to misdiagnosis, delayed treatment, or ethical breaches. Therefore, toxicity management in AI-assisted Pythium diagnosis focuses on risk mitigation, human oversight, and system calibration.[75]

Table. Algorithmic and Diagnostic Risk of AI-Assisted Pythium Diagnosis

Human–Machine Interaction Hazards

• Automation bias: Clinicians may over-trust AI output and skip confirmatory tests.
  • Mandate dual verification (AI + microbiology) and document reasoning if following AI alone.
• Alert fatigue: Excessive probability alerts can desensitize users.
  • Configure tiered alerts (eg, high-probability >0.7 = red; moderate 0.4–0.69 = amber).
• Misinterpretation of heatmaps: Grad-CAM visualizations may be misunderstood as “proof.”
  • Provide educational overlay and disclaimers such as “Regions indicate algorithm attention, not confirmed pathology.”[69]

Table. Data and Privacy Risks with AI-Assisted Pythium Diagnosis 

Hardware/Software-Related Risks

• Edge-device failures: Crashes or misalignment during slit-lamp capture lead to data loss or blurred images.
  • Implement auto-retry capture, local caching, and quality-score rejection.
• Latency in rural networks: Delays in remote inference lead to delayed triage.
  • Use on-device PWA models with offline inference and later synchronization.
• Color imbalance between cameras: Alters AI confidence.
  • Standardize white-balance settings; apply device-specific calibration filters.

Clinical Workflow Considerations

1. Delayed escalation: If AI returns low probability (<0.4) despite classical clinical features, clinicians may delay TPK.
• Implement a “clinical override clause” in which human judgment supersedes AI output.
2. Inconsistent reporting: Missing documentation of AI version/probability may hinder traceability.
• Include AI probability, model version, and timestamp in the patient record.
3. Liability ambiguity: Misdiagnosis from an AI suggestion.
• Define shared responsibility so that AI is clearly an assistive tool and physicians are the final decision-makers.[76]

Table. Post-Deployment Risk Monitoring for AI-Assisted Pythium Diagnosis

 Ethical and Psychological Toxicities

• Patient anxiety: Misunderstanding “AI detected Pythium” as a definitive diagnosis may create emotional distress.
  • Use clear communication scripts: “AI suggests possibility; confirmation pending lab tests.”
• Clinician deskilling: Overreliance on AI pattern recognition.
  • Integrate periodic manual skill assessments and continuing medical education (CME) workshops for corneal image interpretation.
• Equity concerns: Rural/low-resource sites lacking access to AI may lag in early diagnosis.
  • Encourage lightweight, open-source models and publicly shared datasets.[77]

Overall Toxicity Management Framework

Detection → Reporting → Correction → Prevention

1. Detection: Identify misclassification or workflow errors (flagged by the user or an audit).
2. Reporting: Log event in centralized AI-incident registry with anonymized case details.
3. Correction: Retrain model on misclassified cases, update thresholds, or fix capture protocol.
4. Prevention: Continuous education, bias audits, and transparent documentation.[78]

Future Safety Directions

• Explainable AI (XAI) dashboards: Show confidence intervals and rationale.
• Federated monitoring: Share real-time safety metrics across centers.
• Integration with the electronic health record (EHR): Automatically prompt microbiology confirmation orders when AI flags probable Pythium.
• Regulatory pathways: Adhere to WHO AI Ethics 2021, CDSCO (India), and FDA (US) Software as a Medical Device (SaMD) guidance for ophthalmic AI.

Summary

The concept of "toxicity" in the context of AI-assisted Pythium diagnosis is non-biological but high-impact, encompassing diagnostic, operational, and ethical domains. Safe implementation demands a human-in-the-loop framework, robust data governance, and continuous model validation. AI should enhance, not replace, clinical expertise, thereby minimizing both diagnostic error and systemic toxicity.[78]

Staging

While Pythium insidiosum keratitis (PIK) traditionally lacks a universally accepted staging system, as in bacterial or fungal keratitis, a clinico-morphologic staging framework has emerged to guide AI-assisted diagnosis, therapeutic planning, and prognostic stratification. AI and deep learning tools can quantify ulcer morphology, depth, and extent, enabling objective stage classification that correlates with disease severity, perforation risk, and the need for early therapeutic penetrating keratoplasty (TPK).

Table. Clinical–Morphologic Staging System for Pythium Keratitis

Modern AI and imaging workflows enable objective, measurement-based staging for research and teleophthalmology, using the following parameters.

Table. Proposed Imaging-Based (AI-Enabled) Substaging Proposal for PIK

Table. Pathological Staging of PIK (Post-TPK Button/Histopathology)

Integration of AI and Deep Learning into Staging

• Automated segmentation models (U-Net, DeepLab) can outline the infiltrate area and compute the disease index utilizing the established formula (ie, infiltrate area / total corneal area × depth weighting).
• Temporal tracking: Deep learning can quantify daily changes (eg, >15% area expansion in 24 hours predicts progression to Stage III).
• Decision thresholds: Flag for early TPK when AI-predicted Pythium probability ≥0.7 and area growth ≥10%/day.
• Validation: Models trained on serial slit-lamp and IVCM datasets from India and Thailand show a strong correlation (r > 0.85) between AI stage and clinician-assigned grade.[79]

Clinical Utility of Staging

• Research standardization: Enables uniform grading for AI training and outcome comparisons across centers.
• Triage prioritization: AI-derived stage dictates urgency—Stage I–II for conservative management; Stage III–IV for surgical intervention.
• Prognostic counseling: Higher stages are associated with higher rates of graft failure and recurrence.
• Training dataset labeling: Annotated stages enhance the robustness and interpretability of AI models.[80]

Gurnani–Kaur Staging System

The proposed Gurnani–Kaur Staging System (2024) offers a structured, clinically validated, and AI-integrated framework for grading the severity of Pythium keratitis. It is treatment-oriented, emphasizing early recognition, stage-specific therapy, and integration of digital/AI diagnostic support to ensure uniformity in clinical research and teleophthalmology.

Table. Proposed Gurnani–Kaur Staging System for Pythium insidiosum Keratitis

Key Principles of the Proposed Gurnani–Kaur Staging

• Therapy-linked stratification:
  • Stages I–II → Medical management
  • Stages III–V → Surgical intervention ± adjunctive therapy
• Dynamic staging:
  • Incorporates disease progression rate and AI-derived lesion expansion (>10–15%/day) as triggers for upstaging
• AI-integrated quantification:
  • Staging aligns with deep-learning models that measure infiltrate area, depth (via OCT/IVCM), and heatmap intensity to assign probability-based stages
• Surgical reproducibility:
  • TPK margin: 1–1.5 mm beyond visible lesion
  • Donor graft: ≥0.5–1.0 mm oversized
  • Graft size: ≥8 mm preferred for peripheral or limbal disease
• Outcome correlation:
  • Stage I–II: >80% anatomical and visual success with medical therapy
  • Stage III: around 60% globe salvage with early TPK
  • Stage IV–V: guarded prognosis; recurrence risk remains high (10–20%)[81]

The Proposed Gurnani–Kaur Staging System bridges clinical, microbiological, and AI-based diagnostic domains for Pythium insidiosum keratitis. It enables standardized reporting, early escalation to TPK, and real-time AI-guided triage, representing a paradigm shift in how Pythium keratitis is graded and managed in both tertiary and teleophthalmology settings. The proposed Pythium Keratitis Staging Framework, supported by AI and imaging metrics, stratifies disease into four progressive stages—from localized epithelial disease to full-thickness perforation. This approach provides a standardized platform for early identification, appropriate therapy selection, and surgical timing, while enabling AI models to learn reproducible patterns of severity.[15]

Prognosis

The prognosis for Pythium insidiosum keratitis hinges on the timing of diagnosis and whether timely anti-oomycete therapy and early TPK (when indicated) are instituted. Delays of even 48 to 72 hours after clear clinical progression markedly worsen anatomic and visual outcomes.

Expected Outcomes

Typical tertiary-center experience

• Stage I–II (localized/progressive without limbus):
  • Many eyes can be salvaged medically with topical linezolid 0.2% + azithromycin 1% (± oral agents).
  • Anatomic salvage: ~70–85%
  • Functional vision (≥20/200): ~40–60% after optical rehab.
• Stage III (deep/limbal disease):
  • Early TPK is usually required.
  • Anatomic salvage: ~55–70% (higher with large, disease-free margins).
  • Functional vision: ~25–40% (often limited by graft clarity/astigmatism).
• Stage IV (perforation/scleral extension):
  • Guarded; high risk of recurrence and secondary glaucoma/endophthalmitis.
  • Repeat TPK or evisceration may be necessary.
• Recurrence after TPK (Stage V):
  • Typically presents at the graft–host junction within 2–3 weeks.
  • Repeat, larger-diameter TPK improves control; prognosis depends on the speed of re-excision.

Where available, AI-assisted triage (slit-lamp ± IVCM models) can shorten the time to accurate diagnosis and reduce futile antifungal exposure, translating into higher early-TPK rates when appropriate and improved globe salvage.[82]

Table. Adverse Prognostic Indicators in Pythium Keratitis

Causes of Visual Loss Even With Anatomic Salvage

• Graft scarring or vascularization, high irregular astigmatism
• Secondary glaucoma or cataract
• Recurrence at the graft edge requiring repeat TPK
• Macular comorbidity or amblyopia (children)[83]

Counseling and Follow-Up

• Set expectations early; emphasize the possibility of surgery and the need for repeat surgery.
• Schedule daily reviews initially; photograph and measure edges to detect progression.
• Warn about signs of recurrence (eg, new gray “edge,” increased pain, hypopyon).
• After quiescence, plan optical rehabilitation (suture optimization, RGP/scleral lens, later optical keratoplasty if stable).[84]

How Staging Informs Prognosis (Gurnani–Kaur)

• Stage I–II: Generally ffavorablewith medical therapy if progression halts in 48 to 72 hours
• Stage III: Guarded, but early large-margin TPK yields good anatomic outcomes
• Stage IV–V: Poor; prioritize infection control and globe preservation; vision often secondary.

To summarize, early recognition, Pythium-active therapy, and timely TPK (when indicated) with adequate margins are the decisive determinants of outcome; delays and limbal/deep involvement sharply reduce both globe salvage and final vision.[85]

Complications

Pythium keratitis is notorious for its rapid progression, resistance to antifungal therapy, and high complication rates, especially when diagnosis or surgical intervention is delayed. The organism’s unique oomycete cell wall (cellulose, not chitin) and aggressive stromal invasion cause distinctive, often catastrophic sequelae affecting both globe integrity and visual function.[8]

Table. Ocular Complications of Pythium Keratitis

Table. Post-Therapeutic & Surgical Complications of Pythium Keratitis

Table. Diagnostic and AI-Related Complications in Pythium Keratitis Management

Vision-Related Sequelae of Pythium Keratitis

• Permanent corneal opacity/scarring
  • Even with anatomic success, visual acuity is often limited to 20/200 to 20/400 without optical keratoplasty.
• Irregular astigmatism and anisometropia
  • Common after large TPK and may require rigid gas-permeable or scleral lenses.
• Graft vascularization
  • Increases risk of immune rejection and recurrent infection
• Phthisis bulbi/globe loss
  • Seen in up to 10–15% of advanced cases (Stage IV–V) despite surgical intervention [86]

Table. Long-Term Complications of Pythium Keratitis

To summarize, complications of Pythium keratitis span the full spectrum from rapid corneal melt and perforation to post-TPK recurrence and graft failure. The best prevention lies in timely recognition, Pythium-specific therapy, and well-coordinated medical-surgical management. Integration of AI-assisted early diagnosis, adherence to Gurnani–Kaur staging, and multidisciplinary follow-up substantially reduces complication rates and improves long-term globe preservation.[87]