Rethinking Multiple Instance Learning for Corneal Perforation Detection on Radial Anterior Segment Optical Coherence Tomography in Microbial Keratitis

Lucia H Rhode, Kamini Reddy, Folahan Ibukun, Subeesh Kuyyadiyil, Elesh Jain, Gautam Parmar, Rama Chellappa, Nakul S. Shekhawat

Keywords: Anterior Segment Optical Coherence Tomography, Microbial Keratitis, Corneal Perforation, Corneal Imaging, Multiple Instance Learning, Attention Pooling, Weakley Supervised Learning

TL;DR: In small, fixed AS-OCT scan sets for microbial keratitis, simple MIL pooling is enough and performance is driven mainly by encoder quality, not attention-based aggregation.

Registration Requirement: Yes

Abstract: Multiple instance learning (MIL) is a dominant paradigm for weakly supervised pathology detection in medical imaging, yet its assumptions of large, unordered, and heterogeneous bags are violated in several clinical modalities. Radial anterior segment OCT (ASOCT) produces a fixed, small, and anatomically structured bag of six scans at predefined angles, raising the question of whether expressive attention-based pooling provides meaningful benefit under these constraints. We present the first systematic study of MIL pooling for microbial keratitis perforation detection in ASOCT, evaluating mean, max, attention-based MIL (ABMIL), and gated attention pooling across ResNet-50 and ViT-B/16 encoders on an infected-only cohort of 150 eyes (24 perforated) with patient-grouped stratified 5-fold cross-validation. Encoder choice consistently dominated performance, with pretrained ViT feature extractors consistently outperforming convolutional features across all pooling methods. Critically, learned attention pooling yields no meaningful advantage over mean pooling within this small, fixed-bag setting. These findings challenge the prevailing assumption that increasingly expressive MIL pooling is universally beneficial, demonstrating instead that for structured, low-cardinality bags common in certain biomedical imaging modalities, representation quality is the primary driver of performance while complex pooling provides limited gains.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 123