Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2026 Agenda and Abstracts | Next >

2026 OMIG Abstract

Corneal Infiltrate Segmentation of ASOCT Images in Microbial Keratitis via Annotation-Efficient, Domain-Matched Self-Supervised Pretraining

Lucia Rhode¹, Kamini N Reddy², Folahan Ibukun², Subeesh Kuyyadiyil³, Elesh Jain³, Gautam Parmar³, Rama Chellappa¹, Nakul Shekhawat²

¹Whiting School of Engineering, Johns Hopkins University, Baltimore, Maryland; ²Wilmer Eye Institute, Johns Hopkins University, Baltimore, Maryland; ³SNC Chitrakoot, Madhya Pradesh, India

Purpose: Quantitative measurement of infectious stromal infiltrates on anterior segment optical coherence tomography (ASOCT) scans could enable objective monitoring of microbial keratitis (MK) severity and treatment response. However, manual infiltrate annotation on ASOCT scans is time-intensive and expert-dependent, while infiltrate boundaries are inherently ambiguous. This limits the feasibility of large-scale infiltrate segmentation for clinical care or research and necessitates development and validation of automated approaches. Self-supervised pretraining, which learns image structure from unlabeled data, may reduce the number of expert-labeled scans required for training segmentation models. We compared segmentation performance of a model pretrained on unlabeled ASOCT scans from eyes with MK against standard natural-image initialization.

Methods: At each eye-visit, 6 radial ASOCT B-scans centered on the infiltrate were obtained using the Heidelberg Anterion ASOCT platform. An expert grader outlined the hyperreflective stromal lesions corresponding to the infiltrate on each B-scan. A U-Net with a ResNet-34 encoder was trained under two initializations: (1) standard ImageNet-pretrained weights, and (2) masked-autoencoder pretraining on approximately 11,000 unlabeled ASOCT B-scans from eyes with microbiologically confirmed MK. Both were fine-tuned on the identical set of 88 expert-segmented B-scans and evaluated on a held-out test set of 98 radial B-scans from 6 eyes, with each of the test eyes imaged at three sequential disease stages (active infectious infiltrate, mid-healed, post-infectious scar). Test eyes were disjoint from training and pretraining eyes. Metrics included ensemble Dice and 95th-percentile Hausdorff distance (HD95, mm), with per-fold means and standard deviations reported to assess training stability.

Results: Masked pretraining outperformed ImageNet initialization on every metric, including Dice (0.697 vs. 0.573, +0.124 difference) and HD95 (0.567 vs. 0.850 mm, -33% difference). Performance improvements were seen across all three disease stages, with the domain-pretrained model achieving Dice 0.687 + 0.130 in active infectious infiltrates, 0.704 + 0.106 in mid-healed lesions, and 0.700 + 0.218 in post-infectious scars. Corresponding HD95 was 0.791, 0.505, and 0.377 mm, respectively. Boundary error was largest in active infectious infiltrates and smallest in post-infectious scars. Per-fold Dice variability was low (SD < 0.029), indicating stable training.

Conclusions: Pretraining on unlabeled, in-domain ASOCT data yields a substantial and stage-independent segmentation gain over generic natural-image pretraining, and is achievable with fewer than 100 labeled B-scans. This establishes a performance floor for automated ASOCT infiltrate segmentation and demonstrates that unlabeled ASOCT scans carry recoverable structural information.