Automated Detection of Filamentous Fungal Keratitis on Whole Slide Images of Potassium Hydroxide Smears with Multiple Instance Learning.

PURPOSE

The diagnosis of fungal keratitis using potassium hydroxide (KOH) smears of corneal scrapings enables initiation of the correct antimicrobial therapy at the point-of-care but requires time-consuming manual examination and expertise. This study evaluates the efficacy of a deep learning framework, dual stream multiple instance learning (DSMIL), in automating the analysis of whole slide imaging (WSI) of KOH smears for rapid and accurate detection of fungal infections.

DESIGN

Retrospective observational study.

PARTICIPANTS

Corneal scrapings from 568 patients with suspected fungal keratitis; 51% contained filamentous fungi according to human expert interpretation.

METHODS

Dual stream multiple instance learning was employed to analyze WSI of KOH smears. Due to the extensive size of these images, often exceeding 100 000 pixels, conventional computer vision methods (e.g., convolutional neural networks) are not feasible. Dual stream multiple instance learning segments the WSI into patches for analysis, extracting relevant features from each patch and aggregating these to make a comprehensive slide-level diagnosis while generating heat maps to visualize areas contributing most to the prediction. Fivefold cross-validation was used for training and validation, with a hold-out test set comprising 15% of the total samples.

MAIN OUTCOME MEASURES

Accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), F1 score, positive predictive value (PPV), and negative predictive value (NPV) in distinguishing fungal from nonfungal slides.

RESULTS

Dual stream multiple instance learning demonstrated an overall AUC of 0.88 with an accuracy of 79% and an F1 score of 0.79 in distinguishing fungal from nonfungal slides, with sensitivity of 85%, specificity of 71%, PPV of 80%, and NPV of 79%. For "consensus cases," where 2 human graders agreed on the slide interpretation, the model achieved an accuracy of 85% and an F1 score of 0.85. For "discrepant cases," the accuracy was 71% with an F1 score of 0.71. The generated heatmaps highlighted regions corresponding to fungal elements. Code and models are open-sourced and available at https://github.com/Redd-Cornea-AI/KOH-Smear-DSMIL.

CONCLUSIONS

The DSMIL framework shows significant promise in automating interpretation of KOH smears. Its capability to handle large, high-resolution WSI data and accurately detect fungal infections, while providing visual explanations through heatmaps, could enhance the scalability of KOH smear interpretation, ultimately reducing the global burden of blindness from infectious keratitis.

FINANCIAL DISCLOSURES

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