Applications of Computer Vision for Infectious Keratitis: A Systematic Review

Abstract

Clinical Relevance

Corneal ulcers cause preventable blindness in >2 million individuals annually, primarily affecting low- and middle-income countries. Prompt and accurate pathogen identification is essential for targeted antimicrobial treatment, yet current diagnostic methods are costly and slow and require specialized expertise, limiting accessibility.

Methods

We systematically reviewed literature published from 2017 to 2024, identifying 37 studies that developed or validated artificial intelligence (AI) models for pathogen detection and related classification tasks in infectious keratitis. The studies were analyzed for model types, input modalities, datasets, ground truth determination methods, and validation practices.

Results

Artificial intelligence models demonstrated promising accuracy in pathogen detection using image interpretation techniques. Common limitations included limited generalizability, lack of diverse datasets, absence of multilabeled classification methods, and variability in ground truth standards. Most studies relied on single-center retrospective datasets, limiting applicability in routine clinical practice.

Conclusions

Artificial intelligence shows significant potential to improve pathogen detection in infectious keratitis, enhancing both diagnostic accuracy and accessibility globally. Future research should address identified limitations by increasing dataset diversity, adopting multilabel classification, implementing prospective and multicenter validations, and standardizing ground truth definitions.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Corneal ulcers represent a significant global health concern. They are a leading cause of vision impairment and preventable blindness worldwide, affecting >2 million individuals annually, especially in low- and middle-income countries.^1,2 Early detection and prompt initiation of targeted treatment are crucial for preventing poor visual outcomes. However, the diagnostic process relies heavily on expensive and time-consuming laboratory methods and the availability of trained ophthalmologists, who are often scarce where the disease burden is highest. Even in well-resourced settings, human diagnostic accuracy can be suboptimal. Studies suggest that ophthalmologists and cornea specialists may achieve only about 66.0% to 75.9% accuracy in differentiating between bacterial keratitis (BK) and fungal keratitis (FK) on slit lamp photography.³ Such limitations lead to delays in initiating appropriate therapy, ultimately compromising patient outcomes.

Artificial intelligence (AI) and machine learning (ML) have emerged as transformative tools in ophthalmology.⁴ By leveraging advancements in imaging technology and data analysis, AI/ML models can provide immediate diagnostic insights based on images and patient data. These models offer the potential to increase diagnostic accuracy, reduce time-to-diagnosis, and improve access to specialized care in resource-limited environments. Such improvements are particularly relevant in low- and middle-income countries, in which microbiological cultures can take several days, polymerase chain reaction-based assays may be unavailable, and shortages of experienced ophthalmologists compound diagnostic delays.^5,6 With AI/ML-based tools, clinicians can potentially initiate targeted therapy sooner, thereby improving visual outcomes and preventing unnecessary vision loss.

This review aims to provide a comprehensive analysis of current AI/ML-based methods for classifying infectious keratitis into clinically meaningful categories. We evaluate existing AI models, their input modalities, ground truth determination strategies, model architectures, and validation practices. We identify key limitations—such as limited dataset diversity, exclusion of mixed infections, and lack of multilabeled classifications—and propose solutions to enhance generalizability and clinical implementation. By addressing these issues, we hope to clarify the current landscape of AI/ML applications in infectious keratitis diagnosis and guide future research efforts toward scalable, robust, and contextually appropriate models.

Methods

We conducted a systematic review using databases such as PubMed, IEEE Xplore, and Google Scholar, targeting publications available up to the end of December 2024. Our search strategy aimed to capture a broad spectrum of studies at the intersection of AI/ML and keratitis. We employed a combination of keywords and Boolean operators: (“keratitis” OR “infectious keratitis” OR “corneal infection” OR “corneal ulcer”) AND (“AI” OR “artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network”). Reference lists of identified articles were manually screened to find additional relevant studies.

We included peer-reviewed articles published in English that provided empirical data on the development, validation, or application of AI/ML models for diagnosing or managing infectious keratitis, regardless of the study design. Exclusion criteria were applied to articles that did not directly address AI/ML applications in the context of infectious keratitis, were not available in English full text, or were published after the specified cutoff date.

We extracted and analyzed data on model performance, architectures, input types, ground truth determination, dataset representativeness, and model evaluation methodologies. The reviewed papers and their key characteristics are summarized in Table 1, allowing for direct comparison of factors influencing diagnostic performance and generalizability across studies.

Table 1.

Collection and Summary of Reviewed Papers

	Title	Summary	Type of Model	Model Input	Model Output	Training Set	Test Set	Prospective or Retrospective	Method of Ground Truth Determination	Best Architecture	Model Performance	Key Findings
Slit lamp images	Deep sequential feature learning in clinical image classification of infectious keratitis (Xu et al, 20217)	Proposes a sequential-level deep learning model for classifying infectious keratitis using slit lamp images. The model extracts features from image patches in a sequential order from the lesion center outward.	Deep CNN with LSTM	Slit lamp images	Fungal keratitis, bacterial keratitis, herpes simplex virus keratitis, other (corneal dystrophies, phlyctenular keratoconjunctivitis, various corneal tumors, corneal papilloma, corneal degeneration, AK).	1922 images from 747 patients in China	362 images from 120 patients in China	Retrospective	Two-step manual identification by ophthalmologists: (1) image-only diagnosis, (2) diagnosis based on clinical information corroborated by at least 2 of clinical manifestations, treatment response, and pathogen identification	DenseNet + LSTM	80% accuracy on test set, compared to 49.27% by 421 ophthalmologists	Unique use of LSTM to model spatial relationships from lesion center outward.
	From the diagnosis of infectious keratitis to discriminating fungal subtypes: a deep learning-based study (Soleimani et al, 20238)	Proposes 3 CNN models for diagnosing infectious keratitis, differentiating bacterial vs. fungal keratitis, and discriminating filamentous vs. yeast fungal subtypes using slit lamp images.	CNN	Slit lamp images	(1) Infectious keratitis vs. normal, (2) Bacterial vs. fungal keratitis, (3) Filamentous vs. yeast fungi	72% of 9329 images from 977 patients in Iran	20% of images used for fivefold cross-validation	Retrospective	Microbiology culture results	Custom CNNs	Model 1: 99.3% accuracy, AUROC 1.0; Model 2: 84% accuracy, AUROC 0.96; Model 3: 77.5% accuracy, AUROC 0.99	Strong performance in differentiating fungal subtypes.
	Determination of probability of causative pathogen in infectious keratitis using deep learning algorithm of slit lamp images (Koyama et al, 20219)	Developed a deep learning model to determine causative pathogen (bacteria, fungi, Acanthamoeba, HSV) for corneal ulcers using slit lamp images, combining CNN and gradient boosting.	InceptionResNetV2 and Gradient Boosting Decision Tree	Slit lamp images	Bacterial keratitis, fungal keratitis, Acanthamoeba keratitis, HSK. Excluded mixed infections.	1426 slit lamp images collected before March 2019 in Japan	140 images from Japan, 312 web images	Retrospective	Microbiology, PCR, clinical signs, and treatment response. Results were reviewed by 3 independent clinicians who came to unanimous diagnosis.	InceptionResNetV2 + Gradient Boosting	Accuracy: 88%-98%.	Integration of multiple modalities (gradient boosting decision trees with CNN) can improve pathogen classification efficiency.
	Deep learning for discrimination between fungal keratitis and bacterial keratitis: deep keratitis (Ghosh et al, 2022 10)	Developed CNN models using slit lamp images to distinguish fungal keratitis from bacterial keratitis; best performance with an ensemble model.	VGG19, ResNet50, DenseNet121	Slit lamp images	Fungal keratitis versus bacterial keratitis.	1159 BK images, 673 FK images from Thailand	162 BK images, 61 FK images	Retrospective	Microbiology culture or PCR results.	Ensemble model	F1 score: 0.83, AUPRC: 0.904.	CNN with ensemble learning showed the best performance in discriminating FK from BK compared with single architecture tools.
	An image diagnosis algorithm for keratitis based on deep learning (Ji et al, 202211)	Proposes a multitask recognition method using an improved multiattribute network based on ResNet50 to diagnose keratitis from anterior segment images.	Multi-attribute network with ResNet50	Slit lamp images	Five attributes are related to keratitis signs: opacity area, boundary of focus, epithelium integrity, hyperemia, and neovascularization.	668 images from China	286 images from China	Retrospective	Labeled by 3 ophthalmologists.	Improved multi-attribute ResNet50 network	Average accuracy: 84.89%; highest individual attribute accuracy: 89.51%	Multitask learning for detailed symptom recognition.
	Deep learning-based classification of infectious keratitis on slit lamp images (Zhang et al, 202212)	Created a CNN capable of classifying bacterial keratitis, fungal keratitis, Acanthomoeba keratitis, and herpes simplex virus keratitis. Compared results with 3 ophthalmologists.	ResNet, DenseNet/EfficientNet + ResNeXt variations	Slit lamp images	Bacterial keratitis, fungal keratitis, AK, HSK	4347 images from China. Excluded mixed infections.	483 images from China. 200 images from 200 patients (50 from each class) were used for validation set versus ophthalmologist diagnosis.	Retrospective	HSK: clinical history, treatment response, and diagnosis by 3 corneal specialists	KeratitisNet (ResNext101_32 × 16d + DenseNet169)	Accuracy: 77.08,	KeratitisNet demonstrates a good performance on clinical infectious keratitis diagnosis and classification.
									FK, BK, and AK: clinical signs and microbiology (smear or culture) results		AUROC: 0.86 (BK), 0.91 (FK), 0.96 (AK), 0.98 (HSK)
	Deep learning approach in image diagnosis of pseudomonas keratitis (Kuo et al, 2022 13)	Explores using deep learning models to differentiate Pseudomonas keratitis from non-Pseudomonas bacterial keratitis using slit lamp images.	ResNet50, DenseNet121, ResNeXt50, SE-ResNet50, and EfficientNets B0 to B3	Slit lamp images, 224 x 224 pixels.	Pseudomonas keratitis versus non-Pseudomonas keratitis	743 images from Taiwan	186 images from Taiwan.	Retrospective	Microbiology results and 3 clinical expert consensus.	Ensemble 4DL (best combined model), EfficientNet B2 (best single model)	Ensemble results: Accuracy: 72.1%, Sensitivity: 79.6%, Specificity: 57.2%	No statistical difference between AUROC and accuracy among the single or ensemble models. The DL approach could enhance differentiation of Pseudomonas from other bacterial keratitis, but the enhancement may be limited.
	Using slit lamp images for deep learning-based identification of bacterial and fungal keratitis (Hung et al, 202114)	Developed a deep learning model to crop slit lamp images and to distinguish bacterial keratitis from fungal keratitis.	DenseNet variants, EfficientNetB3, InceptionV3, ResNet variants	Cropped slit lamp images	Bacterial keratitis versus fungal keratitis	562 BK images, 342 FK images, from 388 patients from Taiwan	128 BK images, 86 FK images, from 96 patients from Taiwan	Retrospective	Microbiology culture results	DenseNet161	AUROC: 0.85 (BK)	Highlights the use of cropped images for classification. The model has good diagnostic accuracy for BK and FK.
											AUROC: 0.85 (FK)
	Multiscale convolutional neural network for accurate corneal segmentation in early detection of fungal keratitis (Mayya et al, 202115)	Proposes a multiscale CNN for corneal segmentation from slit lamp images, followed by ResNeXt for classifying fungal versus nonfungal keratitis.	Multiscale CNN + ResNeXt	Slit lamp images	Fungal keratitis versus nonfungal keratitis	133 diffuse white light images from the USA and India in addition to 540 public domain images.	133 diffuse white light images	Retrospective	Microbial culture results and labels from 2 ophthalmologists.	Multiscale CNN + ResNeXt	Accuracy: 88.96%	Integrates segmentation and classification for early detection of corneal lesions and detection of FK.
	Deep learning for multitype infectious keratitis diagnosis: A nationwide, cross- sectional, multicenter study (Li et al, 2024 16)	Presents a comprehensive deep learning model for diagnosing various types of infectious keratitis.	DenseNet121, InceptionResNetV2, and Swin Transformer	Slit lamp images	Classification into bacterial, fungal, viral, amebic, and non-infectious keratitis.	10 592 images from 6300 patients in China, including 1590 images of BK, 1728 of FK, 3786 of viral keratitis, 312 of amebic keratitis, and 3176 of noninfectious keratitis.	External dataset of 7206 images from 11 other clinical centers (1548 images of BKs, 1865 of FK, 2188 of viral keratitis, 104 of amebic keratitis, and 1501 of noninfectious keratitis) in addition to a prospective dataset from the training data center of 5257 images (256 of BK, 735 of FK, 2250 of viral keratitis, 30 of amebic keratitis, and 1986 of noninfectious keratitis)	Retrospective and prospective	Microbiology results or typical clinical characteristics, both confirmed by complete response to definitive therapy.	DeepIK (DenseNet variant)	AUROCs of >0.96 for all internal, external, and prospective datasets.	Robust performance with AUROCs >0.96 across internal, external, and prospective datasets
	Deep learning-based classification system of bacterial keratitis and fungal keratitis using anterior segment images (Won et al, 2023 17)	Developed a ResNet-50 classifier with 2 modules: LGM for lesion localization and MAM for handling slit beam artifacts. Improves diagnostic accuracy and interpretability.	ResNet-50 (Baseline), ResNet-50 + LGM, ResNet-50 + MAM, ResNet-50 + LGM + MAM	Anterior segment images (broad-beam and slit beam)	Classification: BK vs. FK	594 images from 88 patients (361 BK, 233 FK) at Samsung Medical Center in the Republic of Korea.	90 images from 19 patients (46 BK, 24 FK) + 98 external images from Google image search and ophthalmology textbooks (open source set)	Retrospective	Corneal scraping and culture	ResNet 50 + LGM + MAM	AUROC (Test set): 0.89; AUROC (External validation): 0.69. Accuracy (Test set): 87.8%; Accuracy (External validation): 71.4%. IOU Localization (ResNet-50 Grad-CAM): 0.175; (ResNet-50 + LGM): 0.489	LGM significantly improves lesion localization. MAM improves robustness to slit beam artifacts. Combined modules enhance diagnostic accuracy over the baseline.
	Advances in the diagnosis of herpes simplex stromal necrotizing keratitis: A feasibility study on deep learning approach (Natarajan et al, 2022 18)	Explores the use of DenseNet-201 and other CNN architectures to differentiate HSV stromal necrotizing keratitis from NVK using slit lamp photographs.	DenseNet-201, ResNet, Inception	Diffusely illuminated slit lamp photographs	HSVNK vs. NVK	267 images (157 HSVNK, 110 NVK) in India	40 images (20 HSVNK, 20 NVK)	Retrospective	PCR-proven active HSVNK and culture-proven NVK	DenseNet-201	Accuracy = 72%, Sensitivity = 69.6%, Specificity = 76.5%, AUROC = 0.73	Demonstrates feasibility of using AI for HSVNK diagnosis in resource-limited settings. DenseNet-201 outperformed ResNet (50% accuracy) and inception (62.5% accuracy).
	Comparisons of artificial intelligence algorithms in automatic segmentation for fungal keratitis diagnosis by anterior segment images (Li et al, 2023 19)	Compared 2 AI models for fungal keratitis diagnosis using anterior segment images. Both models utilized ensemble learning with early fusion, but model 2 had segmented inputs rather than the full image.	Model 1 = DenseNet-121, MobileNetV2, SqueezeNet1_0 + LASSO + MLP Model 2 = Model 1 + Segmentation with manual correction	Slit lamp images	FK vs. non-FK	338 images (168 fungal, 170 nonfungal) in China	85 images (42 fungal, 43 nonfungal)	Retrospective	Corneal scraping with 10% KOH, clinical features validated by senior ophthalmologists.	Model 2	Model 1: AUROC = 0.839, Accuracy = 77.65%, Sensitivity = 86.05%, Specificity = 76.19%. Model 2: AUROC = 0.925, Accuracy = 84.52%, Sensitivity = 90.48%, Specificity = 85.71%	Model 2 demonstrated significantly better performance due to segmentation of lesion areas.
	A deep learning approach in diagnosing fungal keratitis based on corneal photographs (Kuo et al, 2020 20)	Developed a DenseNet-based deep learning model for diagnosing fungal keratitis using corneal photographs. Compared performance against noncorneal and corneal specialists.	DenseNet	Corneal photographs by a slit lamp–mounted camera using white light illumination (no slit beam enhancement)	FK vs. non-FK	288 images (114 FK, 174 non-FK (bacterial, herpes, and parasitic) from Kaohsiung Chang Gung Memorial Hospital, Taiwan	Fivefold cross-validation	Retrospective	Laboratory confirmation using culture, microscopy, and molecular techniques	DenseNet with Grad-CAM++	Sensitivity = 71%, Specificity = 68.4%, Accuracy = 69.4%, AUROC = 0.65	Demonstrated potential for aiding first-line practitioners in rural areas with higher sensitivity than non-corneal specialists. Highlighted challenges in improving specificity.
	Class-Aware Attention Network for infectious keratitis diagnosis using corneal photographs (Li et al, 2022 21)	Proposed a class-aware attention network (CAA-Net) for multiclass classification of infectious keratitis (normal, VK, FK, and BK) using corneal photographs. The CAA-Net is a ResNet34 with several modules incorporated to enhance class-dependent feature learning.	CAA-Net (ResNet34 + CAFL + CACF + CASAF) CAFL = class aware feature learning CASAF = class-aware spatial attention fusion CACF = class-aware context fusion	Slit lamp photographs	Normal, VK, FK, BK	1886 images from 519 patients in the Eye Center of the Second Affiliated Hospital of Medical College of Zhejiang University, China.	Fivefold cross-validation with patient-based split	Retrospective	Corneal scraping, culture, OCT, confocal microscopy, sodium fluorescein staining	CAA-Net with Class-Aware Attention Fusion	Accuracy = 70.17%, Recall = 66.14%, Precision = 66.91%, AUROC (Normal: 0.99, VK: 0.81, FK: 0.82, BK: 0.75)	Enhanced classification performance compared to baseline ResNet34 by incorporating class-attention modules with a custom loss function that amplify input features relevant to class membership.
	Automatic Diagnosis of Infectious Keratitis Based on Slit Lamp Image Analysis (Hu et al, 2023 22)	Proposed a deep learning system using slit lamp images for diagnosing infectious keratitis (normal, VK, FK, and BK). Compared performance of six DL models with ophthalmologists.	EfficientNetV2-M, VGG16, ResNet34, InceptionV4, DenseNet121, ViT-Base	Slit lamp images	Normal, VK, FK, BK	2757 images from 744 patients. 1925 images for training, 301 images for validation. Images taken from the Eye Center at the Second Affiliated Hospital of the Zhejiang University School of Medicine in China.	531 images	Retrospective	Corneal scraping results (culture, PCR, and confocal microscopy) with clinical manifestations and response to respective treatment	EfficientNetV2-M	Accuracy = 73.5%, Recall = 68.0%, Specificity = 90.4%, AUROC = 0.85 (macro-average over all classes)	Demonstrated higher performance (higher AUROC) than 2 ophthalmologists. Heatmaps provided interpretable lesion localization. Better AUROC for VK and FK; challenges with BK classification due to similarities with FK and limited dataset.
	Open-Source Automatic Segmentation of Ocular Structures and Biomarkers of Microbial Keratitis on Slit Lamp Photography Images Using Deep Learning (Loo et al, 2021 23)	Proposes SLIT-Net, a fully automatic deep learning-based segmentation algorithm for identifying ocular structures and MK biomarkers on slit lamp photography images, including stromal infiltrates, hypopyons, WBC borders, and corneal edema.	Modified Mask R-CNN (SLIT-Net)	Diffuse white light and diffuse blue light slit lamp images	Segmentation of pathological (e.g., stromal infiltrates) and nonpathological (e.g., corneal limbus) ROIs	133 eyes (USA and India, 2016-2018)	Sevenfold cross-validation	Prospective	Manual annotation by expert physicians; inter-reader variability analysis between P1 and P2	SLIT-Net with ResNet-50 + FPN backbone	Dice similarity coefficient (DSC) range: 0.62–0.95	Achieves high segmentation accuracy for MK biomarkers, outperforms baseline models like U-Net and Mask R-CNN. Open-source implementation.
Smartphone images	Feasibility assessment of infectious keratitis depicted on slit lamp and smartphone photographs using deep learning (Wang et al, 202124)	Developed deep learning models for classifying slit lamp and smartphone images into 4 classes. Found that global images had slightly higher accuracy than regional images. Compared slit lamp vs. smartphone images.	Deep CNN (ResNet50, DenseNet121, Inception V3)	Slit lamp and smartphone images of the eye/cornea collected in China	Normal, bacterial keratitis, fungal keratitis, herpes simplex virus keratitis	4531 slit lamp images from China	571 slit lamp images and 400 smartphone images	Retrospective	Slit lamp imaging, confocal microscopy, and cornea scraping	InceptionV3	Accuracy (test set): 76.05%, AUROC (slit lamp global images): 0.9588, AUROC (slit lamp regional images): 0.9425, AUROC (smartphone images): 0.8529	Slit lamp images outperformed smartphone images; global image context improves accuracy. Suggested diagnostic value in noncorneal structures (e.g., lids and lashes).
Handheld camera images	Image-based differentiation of bacterial and fungal keratitis using deep convolutional neural networks (Redd et al, 202225)	CNNs trained on handheld camera photographs compared to 12 experts. CNN ensemble achieved a higher AUROC than human ensemble.	MobileNetV2, DenseNet201, ResNet152V2, VGG19, Xception	Handheld camera images of the external eye	FK vs. BK	396 images (215 BK images and 181 FK images) from India (Madurai)	100 images from India (Coimbatore), 80 images from multicenter database Validation: 25 BK and 25 FK images from Coimbatore	Retrospective	Microbiology culture	MobileNetV2	AUROC (MobileNetV2): 0.86 (single center), 0.83 (multicenter)	CNNs outperformed experts in classification accuracy for BK and FK.
	Deep convolutional neural networks detect no morphological differences between culture-positive and culture-negative infectious keratitis images (Kogachi et al, 202326)	Assessed whether CNNs could differentiate between culture-positive and culture-negative corneal ulcer images. CNNs did not reliably predict culture positivity, suggesting no morphological differences.	DenseNet201, MobileNetV2 (pretrained on ImageNet).	Handheld camera images	Culture positive or negative	1578 images (69% culture positive) from India	203 images (70% culture positive) from India	Prospective	Microbiology (culture and smear) results	None	No better than random chance at predicting culture positivity.	No morphological differences exist between culture-positive and culture-negative ulcer images.
	Differentiation of active corneal infections from healed scars using deep learning (Tiwari et al, 202227)	Created a CNN to differentiate active corneal ulcers from healed scars using external photographs.	VGG16 pretrained on ImageNet	Handheld (DSLR) camera images of the external eye	Active corneal infections vs. healed scars	1313 corneal ulcers and 1132 corneal scars from 612 patients from SCUT (USA) and MUTT (India)	Tested on 200 patients in India and 101 images from patients at Stanford University.	Retrospective	Microbiology (culture) results or clinical interpretation by a cornea specialist.	VGG16	India, F1: 92%, Sensitivity: 93.5%, Specificity: 84.42%, AUROC: 0.9731. Stanford, F1: 84.3%, Sensitivity: 78.2%, Specificity: 91.3%, AUROC: 0.9474	Effective differentiation between active infections and scars. High accuracy when used on population outside of training dataset.
	Assessing the impact of image quality on deep learning classification of infectious keratitis (Hanif et al, 202328)	Evaluated CNN for BK and FK using ulcer photographs with varying quality parameters. CNN showed expert-level performance regardless of image quality.	CNN	Handheld camera images of the external eye	BK vs. FK	438 images from 164 patients in India	Not applicable	Retrospective	Microbiology (culture or smear) results	-	-	Demonstrated model robustness across varying image qualities.
In vivo confocal microscopy images	Interpretable deep learning for diagnosis of fungal and Acanthamoeba keratitis using in vivo confocal microscopy images (Essalat et al, 202329)	Introduces a dataset of 4001 IVCM images across 4 classes: fungal keratitis, AK, nonspecific keratitis, and normal corneas. Models use saliency maps to highlight diagnostic regions.	CNN	IVCM images of the cornea	Fungal keratitis, AK, nonspecific keratitis, normal corneas	3000 images from the Central Eye Bank of Iran	1001 images from the Central Eye Bank of Iran	Retrospective	Microbiologic confirmation with smear or culture corroborated with clinical features	DenseNet161	Accuracy: 93.55%, Precision: 92.52%, Recall: 94.77%, F1 Score: 96.93%	Strong focus on model interpretability with saliency maps.
	A structure-aware convolutional neural network for automatic diagnosis of fungal keratitis with in vivo confocal microscopy images (Liang et al, 202330)	Proposes SACNN, a two-stream CNN combining GoogLeNet and VGGNet, to diagnose fungal keratitis from IVCM images.	Two-stream CNN	IVCM images	Hyphae present vs. absent	7278 images (3862 with hyphae) from China	1455 images (772 with hyphae, 683 without hyphae) from China	Retrospective	Microbial culture results; images labeled by 3 experts	SACNN (proposed model)	Accuracy: 97.73%, Sensitivity: 97.02%, Specificity: 98.54%	Combines structural information for improved diagnosis. These results suggest that the proposed neural network could be a promising computer-aided FK diagnosis solution.
	Automatic diagnosis of fungal keratitis using data augmentation and image fusion with deep convolutional neural network (Liu et al, 202031)	Proposes a CNN framework using data augmentation and image fusion for diagnosing fungal keratitis from corneal confocal microscopy images.	AlexNet, VGGNet	Corneal confocal microscopy images	Fungal keratitis vs. normal cornea.	994 abnormal, 876 normal images from China.	Approximately 9% of original dataset used for testing	Retrospective	Microbial culture results	AlexNet	Accuracy: 99.95%	High accuracy is achieved through data augmentation and image fusion techniques.
	The clinical value of explainable deep learning for diagnosing fungal keratitis using in vivo confocal microscopy images (Xu et al, 202132)	Developed an explainable AI system using ResNet and Grad-CAM to diagnose fungal keratitis and compared ophthalmologist performance with and without AI assistance.	ResNet + Grad-CAM	IVCM images, 384 x 384 pixels	Hyphae present (FK) vs. absent (BK)	2088 IVCM images from China	1089 images from China	Retrospective	Manual labels agreed upon by 3 cornea experts	ResNet with Grad-CAM	AUROC: 0.983, Accuracy: 0.965, Sensitivity: 0.936, Specificity: 0.982	Demonstrates clinical value of explainable AI in assisting diagnosis is greater than AI-assisted or unassisted.
	A deep transfer learning framework for the automated assessment of corneal inflammation on in vivo confocal microscopy images (Xu et al, 202133)	Developed transfer learning models to detect dendritic and inflammatory cells in IVCM images.	VGG16, ResNet101, Inception V3, Xception, Inception-ResNetV2	IVCM images	Detection of activated dendritic cells and inflammatory cells	3453 images from China	558 images (external validation set)	Retrospective	Manual labels by 3 cornea specialists	Inception-ResNetV2	AUROC: 0.96 (dendritic cells), 0.99 (inflammatory cells). Accuracy: 0.93 (dendritic cells), 0.98 (inflammatory cells)	Applies transfer learning for inflammation assessment. The deep transfer learning models provide an automated analysis of corneal inflammatory cellular components with high accuracy.
	Application of image recognition-based automatic hyphae detection in fungal keratitis (Wu et al, 201834)	Proposes an image recognition-based automatic hyphae detection system using IVCM images. Evaluates the sensitivity and specificity of automatic detection for fungal keratitis compared to ophthalmologist detection on IVCM and KOH corneal smear results.	SVM	IVCM images	Detection of fungal hyphae and density quantification	56 fungal keratitis cases and 23 bacterial keratitis cases from Jinan, China	Not explicitly mentioned	Prospective	Fungal and bacterial cultures.	SVM using adaptive robust binary pattern (ARBP) for feature extraction	Sensitivity: 89.29%, Specificity: 95.65%, AUROC: 0.946	Demonstrates high sensitivity and specificity compared to manual methods. Enables quantitative evaluation of hyphae density, aiding less experienced clinicians. Hyphae density index correlates well with symptom severity.
	An artificial intelligence approach to classify pathogenic fungal genera of fungal keratitis using corneal confocal microscopy images (Tang et al, 202335)	Proposes a deep learning model to classify Fusarium and Aspergillus fungal genera using IVCM images. It compares a hybrid deep learning model (Inception-ResNetV20) with a decision-tree classifier compared to a DL model with a fully connected layer classifier.	Inception-ResNetV2	IVCM images	Fusarium vs. non-Fusarium, Aspergillus vs. non-Aspergillus	3030 images for training from China	334 images for testing taken from the training set without patient-level splitting	Retrospective	Microbiological culture results	Inception-ResNetV2 with fully connected layers	Fusarium: AUROC = 0.887, Accuracy = 0.817, F1 score = 0.749; Aspergillus: AUROC = 0.828, Accuracy = 0.757, F1 score = 0.716	Demonstrates non-invasive genus (Fusarium vs non-Fusarium, Aspergillus vs non-Aspergillus) identification of fungal keratitis with high accuracy, a task extremely difficult to perform for humans based on IVCM images. Fully connected layers outperform decision-tree classifiers.
	Deep learning-based automated diagnosis of fungal keratitis with in vivo confocal microscopy images (Lv et al, 202036)	Developed a ResNet-based deep learning model to automatically diagnose fungal keratitis using IVCM images. Evaluates the model’s diagnostic performance for detecting fungal hyphae.	ResNet with 101 layers and 33 blocks	IVCM images	Detection of fungal keratitis (presence of fungal hyphae)	2088 images (688 positive, 1400 negative) from China	535 images (172 positive, 363 negative) from the same center in China	Retrospective	Fungal culture results	ResNet (101-layer)	AUROC: 0.9875, Accuracy: 96.26%, Sensitivity: 91.86%, Specificity: 98.34%	Included cases of corneal tissue with inflammatory cells or activated dendritic cells in the negative class of the test set, which suggests that the model focused on the filaments rather than inflammation.
	Two-stage deep neural network for diagnosing fungal keratitis via in vivo confocal microscopy images (Li et al, 202437)	Developed a two-stage deep learning/multiple instance learning framework for diagnosing FK using IVCM images. Stage 1 detects FK at the individual image level, while stage 2 uses multi-instance learning to process a sequence of neighboring images from a patient and make a prediction.	Two-stage multiple instance learning framework. Backbone: ResNet18 -ResNet 34 – PoolFormer -SwinTransformer Transformer-based aggregator network.	IVCM images	FK detection at individual image, image sequence, and patient levels	59 165 images from 227 patients in China	28 885 images from 111 patients.	Retrospective	Corneal scraping microscopy or positive fungal cultures; validated by multiple ophthalmologists	Two-stage multiple instance learning framework with Swin Transformer in stage 1 and transformer-based network in stage 2	Sensitivity = 94.44%, Specificity = 100%, Accuracy = 96.23%	Multi-instance learning enhances FK detection by leveraging spatial relationships in image sequences, improving specificity, and achieving higher sensitivity than expert ophthalmologists in image-based diagnosis.
	Medical image management and analysis system based on web for fungal keratitis images (Hou et al, 202138)	Developed a web-based medical image management system with integrated computer-aided diagnosis for fungal keratitis. Utilized AlexNet, ZFNet, and VGG16 for classifying fungal keratitis using confocal microscopy images.	CNN: AlexNet, ZFNet, VGG16	Confocal microscopy images	Normal vs. FK	1310 images (696 fungal, 614 normal) from a hospital laboratory with unspecified location, likely in China.	560 images (298 fungal, 262 normal)	Retrospective	Diagnoses by experienced doctors based on confocal microscopy features	VGG16 and weighted average ensemble of CNNs	Best ensemble: Accuracy = 99.64%, Sensitivity = 99.66%, Specificity = 99.62%	Integrating CNNs improves classification performance. Web-based system enhances diagnostic efficiency for fungal keratitis.
Corneal smear	Automated Detection of Filamentous Fungal Keratitis on Whole Slide Images of Potassium Hydroxide Smears with Multiple Instance Learning (Assaf et al, 202439)	Proposes a deep learning model (DSMIL) to analyze KOH smears for fungal keratitis detection. Employs a multi-instance learning approach to address the challenges of WSI and patch label imbalance.	Multiple instance learning	WSIs of corneal KOH smears	Filamentous fungi vs. nonfungal smears	246 fungus-positive eyes and 254 fungus-negative eyes from South India.	41 fungus-positive and 31 fungus-negative eyes.	Retrospective	WSI-level KOH smear reading by 2 doctors with adjudication by a specialist.	ResNet18 as embedder + DSMIL as aggregator	AUROC = 0.88, Accuracy = 79%, Sensitivity = 85%, Specificity = 72%	Highlights the potential of DSMIL for automated fungal keratitis diagnosis using WSI of KOH smears, enabling faster and scalable diagnostics.
Multimodal approaches	A knowledge-enhanced transform-based multimodal classifier for microbial keratitis identification (Wu et al, 2023)40	Multimodal model using CNN and BERT to classify bacterial vs. fungal keratitis, incorporating slit lamp images and medication text data. Excluded mixed infections with viral or Acanthamoeba.	CNN +BERT	Slit lamp images (initial presentation and first follow-up), medication text records	Bacterial keratitis, fungal keratitis. Excluded mixed infections.	558 images from 279 patients in China	146 images from 73 patients in China	Retrospective	Clinical diagnosis or treatment response or pathogen identification	ResNet152 + BERT	93% accuracy on test set using image and text inputs	Novel integration of medication data and images for keratitis classification.
	Multimodal Deep Learning for Differentiating Bacterial and Fungal Keratitis Using Prospective, Representative Data (Prajna et al, 202441)	Compares clinical data, computer vision, and multimodal models for bacterial vs. fungal keratitis differentiation using a prospective, representative dataset from South India.	EfficientNetB7 (Computer Vision), Feedforward Neural Network (Clinical Data), Multimodal Model	Corneal photographs + structured clinical data	BK vs. FK	1047 images from 419 patients (MADURAI dataset, India)	421 images from 180 patients (MADURAI dataset)	Prospective	Microbiological evidence (culture and smear)	EfficientNetB7	AUPRC: 0.94, AUROC: 0.81, Accuracy: 77%, F1 Score: 0.85 (Computer Vision Model)	Prospective representative data improves model generalizability. The computer vision model outperformed both the clinical data model and the multimodal model.
Structured clinical data	Data-driven approach to eye disease classification with machine learning (Malik et al, 201942)	Proposes a framework to standardize clinical eye examination data for ML prediction of eye diseases based on features like age, medical history, and examination findings.	Decision tree, Random forest, Naive Bayes, Neural network	Structured clinical data on symptoms, examination findings, age, medical history	Classification of 52 eye conditions	241 patients in Pakistan	103 patients	Retrospective	Expert physician diagnosis	Random forest	Accuracy: 86.63% on 52 eye diseases	Demonstrates potential of ML on structured clinical data.
	Development and multicenter validation of machine learning model for early detection of fungal keratitis (Wei et al, 202343)	Developed ML models using clinical signs assessed on slit lamp images to diagnose FK, achieving high sensitivity and specificity on a prospective cohort.	Binary logistic regression, random forest, decision tree	12 clinical signs assessed by ophthalmologists	Predicting FK	713 FK cases, 334 non-FK cases from China	External validation set: 210 FK and 210 non-FK cases from multicenter prospective cohort	Retrospective development set, prospective validation set	Microbiological (culture/smear) results, clinical response to treatment	Binary logistic regression	AUROC: 0.903, Accuracy: 90.5%, Sensitivity: 90.7%, Specificity: 89.9%	Effective use of ML on clinical signs for early FK detection.

AUPRC = area under the precision–recall curve; AUROC = area under the receiver operating characteristic curve; AK = Acanthamoeba keratitis; BERT = bidirectional encoder representations from transformers; BK = bacterial keratitis; CNN = convolutional neural network; DL = deep learning; DSLR = digital single-lens reflex; DSMIL = dual stream multiple instance learning; FK = fungal keratitis; Grad-CAM = gradient-weighted class activation mapping; HSK = herpes simplex keratitis; HSV = herpes simplex virus; HSVNK = herpes simplex stromal necrotizing keratitis; IVCM = in vivo confocal microscopy; KOH = potassium hydroxide; LGM = lesion guiding module; LSTM = long short-term memory; MAM = mask adjusting module; ML = machine learning; MK = microbial keratitis; MUTT = Mycotic Ulcer Treatment Trial; PCR = polymerase chain reaction; ROI = regions of interest; SACNN = structure-aware convolutional neural network; SCUT = Steroids for Corneal Ulcers Trial; SVM = support vector machine; VGG = Visual Geometry Group; VK = viral keratitis; WSI = whole slide image.

This systematic review adhered to the principles outlined in the Declaration of Helsinki. Ethical approval was not required, as the study synthesized data from previously published studies without direct involvement of human participants.

Results

Study Characteristics and Geographic Diversity

Our literature search identified 37 studies published between 2017 and 2024 that developed or validated AI models for pathogen detection or related classification tasks in infectious keratitis. Among these, 19 studies (51%) included data from China,^{7,11,12,16,19,21,22,24,}30, 31, 32, 33, 34, 35, 36, 37, 38^,40,43 9 (24%) included data from India,^15,18,23,25, 26, 27, 28^,39,41 3 (8%) included data from the United States,^15,23,27 3 (8%) included data from Taiwan,^13,14,20 2 (5%) included data from Iran,^1,2 and 1 study each (3% each) included data from Japan,⁹ Thailand,¹⁰ the Republic of Korea,¹⁷ and Pakistan.⁴² Some (3 studies; 8%) combined data from multiple regions.^15,23,27 Most studies (31, 84%) relied solely on single-center datasets without external validation; 6 (16%) employed prospective validation^{16,23,26,34,41,43}; and 7 (19%) included external validation sets.^{7,9,16,17,25,27,43} Only 4 studies (1%) used or referenced publicly available datasets,^9,15,17,29 underscoring limited data sharing.

The geographic variability in pathogen prevalence (e.g., fungal infections being more common in humid regions like India⁴⁴) highlights the importance of regional data during model training. Models developed with single-region data risk overfitting to local pathogen distributions or site-specific nuances in data acquisition methods, limiting their applicability elsewhere. Redd et al²⁵ trained a convolutional neural network (CNN) on handheld camera images from a single site in South India to avoid label leakage observed during initial multicenter training, where site-specific patterns (e.g., differences in pathogen prevalence) impaired model generalizability. The final model achieved an area under the receiver operating characteristic curve (AUROC) of 0.86 on a single-center test set and 0.83 on a multicenter test set, demonstrating reasonable generalizability across centers despite geographic variability. Federated learning offers an alternative strategy to multicenter training, which provides better data security but has yet to be evaluated for infectious keratitis model training.⁴⁵

Model Tasks

The studies encompassed various classification and segmentation tasks. Eight studies (22%) focused on bacterial versus fungal classification.^{10,14,17,25,28,33,40,41} Ghosh et al¹⁰ used an ensemble of ResNet, DenseNet, and Visual Geometry Group architectures for bacterial versus fungal, reporting F1 = 0.83. Soleimani et al⁸ proposed 3 CNNs for diagnosing keratitis and differentiating bacterial versus fungal versus filamentous subtypes, with 84% accuracy (AUROC 0.96) for BK versus FK. Nine studies (24%) addressed multiple pathogen types (bacterial, fungal, viral, protozoal, or noninfectious).^{7,9,12,16,19,21,22,24,29} Zhang et al¹² classified bacterial, fungal, Acanthamoeba, and herpes simplex keratitis, achieving 77.08% accuracy (AUROCs of 0.86–0.98). Eight studies (22%) targeted fungal detection versus nonfungal categories.^13,15,19,30, 31, 32^,36,37 Twelve studies (32%) performed other tasks, including differentiating active infections from healed scars,²⁷ identifying fungal genera,³⁵ assessing culture positivity,²⁶ or evaluating image quality effects.²⁸ Tiwari et al²⁷ used VGG16, reporting an F1 of 0.92 and an AUROC of 0.9731 for separating active ulcers from scars. None of the studies implemented multilabeled classification for polymicrobial infections. This is problematic considering that up to 32% of infectious keratitis cases may be polymicrobial.⁴⁶

Model Architectures

Convolutional neural networks dominated the landscape, with 34 studies (92%) employing CNN-based frameworks.7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33^,35, 36, 37, 38, 39, 40, 41 Two studies (5%) integrated multiple instance learning,^37,39 and 4 studies (11%) explored transformer-based approaches,^16,37,39,40 including Swin Transformers or bidirectional encoder representations from transformers for multimodal data integration, such as Wu et al⁴⁰ who integrated bidirectional encoder representations from transformers with ResNet for BK versus FK detection on slit lamp images and medication text data. Some studies combined CNNs with gradient boosting⁹ or segmentation models,^15,19,23 whereas a small minority (3 studies; 8%) utilized traditional ML methods^34,42,43 and 1 (3%) incorporated long short-term memory to encode spatial context (Xu et al⁷). The overwhelming emphasis on CNNs reflects their suitability for image-based tasks, although novel architectures may expand future capabilities.

Model Inputs

Model inputs were diverse. Seventeen studies (46%) used slit lamp images,7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 10 (27%) used in vivo confocal microscopy (IVCM) images,29, 30, 31, 32, 33, 34, 35, 36, 37, 38 4 (11%) used handheld camera images,25, 26, 27, 28 1 study (3%) focused on smartphone images,²⁴ 1 study (3%) analyzed corneal smear whole slide imaging,³⁹ 2 studies (5%) adopted multimodal inputs (combining imaging with text data),^40,41 and 2 studies (5%) relied solely on structured clinical data.^42,43 Some studies (3 studies; 8%) incorporated segmentation of corneal lesions,^15,19,23 and a few (2 studies; 5%) explored symptom-level classification tasks.^11,33 Essalat et al²⁹ used IVCM images for four-class classification (fungal, Acanthamoeba, nonspecific, and normal), reporting 93.55% accuracy. Similarly, Assaf et al³⁹ applied a multiple instance learning framework to analyze entire whole slide images of potassium hydroxide smears, achieving an AUROC of 0.88 for fungal versus nonfungal differentiation.

Imaging datasets ranged from small (<500 images) to large (>2000 images), with 15 studies (41%) reporting larger datasets exceeding 2000 images.^{7,8,10,12,16,22,24,27,29,30,32,33,}35, 36, 37 Studies comparing input types, such as slit lamp versus handheld and smartphone images, demonstrated higher diagnostic accuracy for slit lamp images. For instance, Wang et al²⁴ found that slit lamp images had superior diagnostic accuracy (AUROC of 0.96) compared with smartphone camera images (AUROC of 0.85), likely due to better image quality and the ability to localize infiltrate depth within the cornea using slit beam illumination.

Studies comparing input types and preprocessing strategies highlighted the impact of image context on diagnostic accuracy. Wang et al²⁴ also showed that models trained on global, unsegmented images of the eye and adnexa slightly outperformed those trained on regionally cropped images around the limbus (AUROC 0.96 vs. 0.94), suggesting valuable diagnostic information may reside in noncorneal structures like lids, lashes, and sclera and underscoring the importance of incorporating contextual data in model development. However, care must also be taken to avoid incorporating demographic bias into the model by training on nonrepresentative datasets that include nondiagnostic information such as skin pigmentation or other racial or ethnic factors, which may erroneously influence model predictions and reduce generalizability.

Ground Truth Determination

Establishing a robust ground truth for supervised learning proved challenging in many studies. Various reference standards were used: 13 studies (35%) relied on microbiological confirmation alone (based on corneal scraping, smear, culture, or polymerase chain reaction)^{8,10,13,17,18,25,26,28,31,}34, 35, 36^,41; 10 studies (27%) combined expert assessment with microbiological confirmation^{7,9,12,13,15,19,22,27,29,30,37,39}; 5 studies (14%) used expert consensus alone^{11,23,32,33,42}; and 7 studies (19%) incorporated microbiological confirmation plus treatment-response criteria.^{7,9,12,16,22,40,43} Separately, 1 study used IVCM as the sole reference standard³⁸; another combined confocal imaging with slit lamp photographs and cornea scraping²⁴; and a third integrated microbiological testing, confocal microscopy, fluorescein staining, and OCT to establish the final diagnosis.²¹

For example, Zhang et al¹² established ground truth labels for their dataset of 4830 slit lamp images by integrating clinical manifestations, treatment responses, and pathogen identification. Cases of bacterial, fungal, and Acanthamoeba keratitis were confirmed through clinical signs and at least 1 positive laboratory result, such as a smear or microbial culture, whereas herpes simplex keratitis was diagnosed based on clinical history, manifestations, and response to antiviral therapy. Using these criteria, their deep learning model achieved an overall accuracy of 77.08%, with AUROC values ranging from 0.86 for BK to 0.98 for herpes simplex keratitis. In contrast, Soleimani et al⁸ used culture results, reporting 84% accuracy for BK vs. FK on slit lamp images. This heterogeneity in labeling complicates direct performance comparisons between studies. It also underscores the fact that none of these labeling methods have perfect sensitivity or specificity, meaning these models are being trained with noisy labels. This is particularly problematic for supervised learning, which assumes the labels are an accurate representation of the ground truth. Alternative strategies such as weak supervision using soft labels or self-supervised learning may help address this issue for infectious keratitis model development.

Beyond Pathogen Detection

Several studies (6 studies; 16%) investigated applications beyond pure pathogen detection.26, 27, 28^,33, 34, 35 These included distinguishing active infections from healed scars (Tiwari et al²⁷), assessing the impact of image quality (Hanif et al²⁸), automated hyphae detection (Wu et al³⁴), quantifying corneal inflammation (Xu et al³³), and differentiating fungal genera (Tang et al³⁵). Such efforts broaden the scope of AI in infectious keratitis by addressing clinically relevant aspects of disease management. Integration of clinical data into models (2 studies^40,41; 5%) showed potential for improving diagnostic accuracy over imaging-only approaches, although results varied based on data quality and disease complexity.

Model Validation and Performance Evaluation

Performance varied with input quality, dataset size, validation strategy, and ground truth methods. For studies focused strictly on bacterial versus fungal classifications, AUROC values ranged from 0.81⁴¹ to 0.89,¹⁷ and accuracy ranged from 71%¹⁷ to 93%.⁴⁰

Models employing higher-quality imaging and larger datasets reported higher AUROCs, often exceeding 0.90.^{8,16,24,36,41} Studies that included prospective or external validation sets frequently observed decreases in performance metrics compared with internal validation alone.^16,17 Seven studies (19%) compared model performance to ophthalmologists, with all of them reporting superior AI performance.^{7,9,12,22,25,32,37}

Discussion

Our review reveals that AI applications in infectious keratitis diagnosis are rapidly evolving yet face several limitations. These findings indicate that while AI models demonstrate promising performance in pathogen classification and related tasks, the lack of multilabeled classification and the exclusion of mixed infections limit applicability in routine clinical practice. The high reliance on single-institution retrospective datasets and the low frequency of prospective and external validation reduce confidence in model generalizability. Most studies used CNN-based architectures and image-based inputs, but there remains potential for exploring more diverse architectures, multimodal strategies, larger and more geographically diverse datasets, and standardized ground truth determination methods. Furthermore, the inclusion of multimodal data, multi-instance learning, and transformer-based architectures may yield more robust and generalizable models. Finally, a unified approach to ground truth labeling, addressing multiple pathogens concurrently, and testing across diverse populations would likely improve translation into clinical practice.

Although our manuscript was under development, Ong et al⁴⁷ published a valuable meta-analysis of image-based deep-learning classifiers for infectious keratitis through July 2024. Our review extends this landscape by including the most recent publications (through December 2024) and by embracing a wider array of methodologies—segmentation-driven networks, multimodal fusion of images with clinical or textual data, whole-slide smear analysis via multi-instance learning, smartphone-based approaches, and prospective validation designs. At the same time, we have focused our inclusion strictly on organism-specific classification, excluding broader corneal-triage models. Together, these updates offer an up-to-date synthesis of emerging AI strategies in infectious keratitis and highlight avenues for improving model robustness, multilabel pathogen detection, federated learning, and implementation in clinical practice.

Generalizability emerged as a key challenge facing the field. Most studies were conducted in single regions without external validation, raising concerns about model adaptability to different patient populations, pathogen profiles, and image acquisition protocols. Geographic variability in pathogen prevalence suggests that AI models must be trained on diverse datasets, potentially implementing additional models fine-tuned to specific populations with unique disease distributions. Techniques like federated learning, in which models are trained across multiple sites without direct data sharing, could enhance generalizability while preserving data privacy.⁴⁵ Similarly, employing data augmentation and transfer learning, as well as exploring transformer-based and multiple instance learning architectures, may help models be more robust and generalize better. Developing domain-specific foundation models may further improve the overall performance of AI/ML applications for infectious keratitis, as has been demonstrated for computer vision of fundus imaging.⁴⁸

Very few studies included data collected over a full calendar year, potentially limiting their ability to capture seasonal variations in pathogen prevalence and clinical presentation. Incorporating year-round patient recruitment could enhance the robustness of models and their applicability across different seasons. Additionally, more widespread prospective data collection would likely improve model generalizability because prospective cohorts can mitigate biases inherent in retrospective data and better approximate routine clinical conditions.

Ground truth determination remains nonuniform across studies, relying heavily on microbiological cultures that can be slow and resource-intensive and have low sensitivity.49, 50, 51, 52, 53 Integrating multiple diagnostic sources, including clinical judgment, polymerase chain reaction-based assays, confocal microscopy, and treatment response, can produce more robust training datasets. Weakly supervised or unsupervised learning may also help overcome labeling challenges, leveraging large volumes of unlabeled ophthalmic data to extract meaningful features without full reliance on definitive ground truth labels.^54,55

Model inputs strongly influence performance. Slit lamp imaging is widely available, and models trained on these images often achieve higher accuracy. However, the finding that global, unsegmented images outperformed regionally cropped images suggests that noncorneal ocular structures may offer additional diagnostic clues. Incorporating alternative imaging modalities such as IVCM or OCT might further improve performance, although these modalities may not be readily available in all settings. Smartphone imaging, despite variable quality, could democratize access to AI diagnostics in resource-limited areas if models are robust to image variability. Studies like Hanif et al's demonstrated that AI models could maintain diagnostic accuracy even with relatively lower-quality images, suggesting potential for broader application in settings with limited resources.²⁸

Many studies considered only binary or multiclass pathogen classifications, excluding scenarios with multiple concurrent infections. Developing multilabeled classifiers would better reflect clinical reality, where mixed infections are not uncommon.⁴⁶ Addressing this complexity would require larger, more diverse datasets and creative strategies to handle low-prevalence scenarios.

Although some studies integrated structured clinical data or natural language processing of treatment records alongside imaging, such multimodal models remain rare. Combining imaging with relevant clinical metadata (e.g., patient demographics, geographic location, systemic comorbidities, and local climatic conditions) may improve model accuracy and decision-making relevance and offers the added benefit of improved model explainability. Nevertheless, such complexity must be balanced with practicality and cost-effectiveness, especially in low- and middle-income countries where expensive diagnostic tools may be scarce.

From a practical standpoint, implementing AI in the clinic entails ensuring interoperability with health information systems, compliance with regulations, cost-effectiveness, and training end-users to trust and interpret model outputs. Rigorous prospective validation, standardized performance metrics, open data sharing, and transparent reporting will be critical for achieving regulatory approval and clinical acceptance. Interdisciplinary collaborative efforts between clinicians, data scientists, and policymakers will be paramount in ensuring that AI tools meet clinical standards and genuinely improve patient care. Our review found that most datasets remain private, limiting external validation and collaboration opportunities.

In the future, exploring unsupervised or self-supervised learning could tap into unlabeled data and reveal novel biomarkers or pathogen-specific imaging patterns.⁵⁶ Additionally, calibrating AI models so that probability estimates reflect true likelihoods of disease and systematically incorporating feedback from clinicians to iteratively refine models could enhance clinical relevance. Taken together, these advancements will help move AI from promising prototypes to widely implemented solutions, ultimately improving accessibility, diagnostic accuracy, and patient outcomes in the management of infectious keratitis.

Manuscript no. XOPS-D-24-00592.