Abstract
Purpose :
This study aims to address the challenges in diagnosing fungal keratitis in corneal KOH smears, which are currently reliant on time-consuming manual examination with variable expertise. We aim to evaluate the efficacy of Dual Stream Multiple Instance Learning (DSMIL) in automating the analysis of whole slide images (WSI) for rapid and accurate detection of fungal infections, a crucial step in initiating appropriate treatment and reducing unnecessary use of antibiotics or antifungals.
Methods :
We employed DSMIL, a deep learning network, to analyze WSIs of KOH smears. Due to the extensive size of these images, often exceeding 100,000 pixels in resolution, conventional computer vision algorithms like CNNs are not feasible. To address this, DSMIL segments the WSI into patches for analysis, extracting relevant features from each patch and aggregating these to make a comprehensive slide-level diagnosis. The study utilized patient-level labels for training the model, analyzing 388 corneal scrapings, with 260 confirmed as fungus-positive. A hold-out test set comprised 15% of the total samples, and statistical analysis was conducted to assess the model's accuracy.
Results :
Preliminary results demonstrate an accuracy of approximately 80% in distinguishing fungal from non-fungal slides, with an Area Under the Receiver Operating Characteristic (AUROC) of 0.76.
Conclusions :
The study highlights the potential of DSMIL in enhancing the speed and accuracy of fungal infection detection in corneal smears. Adapting to the large-scale challenges posed by whole slide imaging, DSMIL offers significant advancements in ophthalmological diagnostics. Notably, this algorithm could serve as an initial assessment tool in areas lacking specialized physicians, providing crucial, timely insights while awaiting expert evaluation. Such an approach could improve patient outcomes through faster, more precise diagnosis and targeted treatment of fungal keratitis, especially in resource-limited settings.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.