Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2024 Agenda and Abstracts | Next >

2024 OMIG Abstract

Automated Detection of Bacterial Keratitis on Whole Slide Images of Gram Stains Using Dual Stream Multiple Instance Learning

Jad F. Assaf¹, Lalitha Prajna², Ramesh Gunasekaran², Thomas M Lietman³, Jeremy D Keenan³,
J Peter Campbell¹, Xubo Song⁴, Travis K Redd<sup>1

1</sup>Casey Eye Institute, Oregon Health & Science university, Portland, OR; ² Department of Ocular Microbiology, Aravind Eye Hospital and PG Institute of Ophthalmology, Madurai, Tamil Nadu, India;
³ Francis I. Proctor Foundation, University of California San Francisco, CA; ⁴ Department of Medical Informatics and Clinical Epidemiology and Program of Computer Science and Electrical Engineering,
Oregon Health & Science University, Portland, OR

Purpose: This study evaluates the efficacy of Dual Stream Multiple Instance Learning (DSMIL) in automating the analysis of whole slide images (WSI) of Gram stains for diagnosing bacterial keratitis. The research addresses the challenges associated with manual examination of Gram stain smears, which are time-consuming and require expert interpretation.

Methods: A deep learning framework, DSMIL, was employed to analyze WSIs of Gram stain smears. To overcome the computational limitations imposed by the high resolution of these images (often exceeding 100,000 pixels), DSMIL divides the WSI into smaller patches, extracts relevant features, and aggregates them to generate a comprehensive slide-level diagnosis. The study analyzed 906 Gram-stained slides, of which 164 were confirmed bacterial-positive, using patient-level labels for training. A hold-out test set comprising 15% of the total samples was utilized for evaluation.

Results: Preliminary analysis indicates that the DSMIL approach achieved an accuracy of approximately 76% in differentiating bacterial from non-bacterial slides.

Conclusions: This investigation demonstrates the potential of DSMIL in enhancing the efficiency and accuracy of bacterial infection detection in Gram stain smears. By addressing the challenges associated with large-scale whole slide imaging, DSMIL offers significant advancements in automated diagnostic processes. This algorithm may prove particularly valuable in resource-limited settings with limited access to trained experts , providing rapid, actionable insights. The implementation of this approach could potentially improve patient outcomes through expedited, more precise diagnoses and targeted treatment of bacterial infections.


Disclosure:

N (LP, RG, TML, JDK, JPC, XS, TKR); O (JFA, NeuralVision - FZCO, Dubai, UAE)