Ocular Microbiology and Immunology Group
Back to OMIG Main Page

< Previous | 2025 Agenda and Abstracts | Next >

2025 OMIG Abstract

Real World Deployment of an AI Model for Potassium Hydroxide Smear Interpretation Reveals Critical Implementation Barriers

Jad F. Assaf¹, Hady Yazbeck¹, Phit Upaphong¹, John Jackson¹, Prajna N. Venkatesh², Lalitha Prajna³, Ramesh Gunasekaran³, Karpagam Rajarathinam³, Thomas M Lietman⁴, Jeremy D Keenan⁴,
J Peter Campbell¹, Xubo Song⁵, Travis K Redd<sup>1*

1</sup>Casey Eye Institute, Oregon Health & Science University, Portland, Oregon; ²Department of Cornea and Refractive Surgery, Aravind Eye Hospital, Madurai, India; ³Department of Ocular Microbiology, Aravind Eye Hospital and PG Institute of Ophthalmology, Madurai, Tamil Nadu, India; ⁴Francis I. Proctor Foundation, University of California San Francisco, California; ⁵Division of Oncological Sciences, Oregon Health & Science University, Portland, Oregon

Purpose: To evaluate the field performance of a previously validated dual stream multiple instance learning network (79% slide level accuracy during development in Madurai, India) for detecting filamentous fungus on potassium hydroxide (KOH) corneal smears, and to identify real world factors responsible for performance degradation.

Methods: Prospective, multicenter implementation study (May–July 2025) at three Aravind Eye Hospital sites (Coimbatore, Pondicherry, Madurai). The trained model was integrated into manual whole slide imaging (WSI) workstations running Microvisioneer MvSlide. Corneal scrape smears were digitized at 20x magnification. WSI-based AI predictions (fungus +/–) were compared with reference standard light microscopy performed through binoculars. The primary endpoint was slide level accuracy. Secondary analyses assessed operator level accuracy and the proportion of human interpretations revised after reviewing the AI classification and its probability heat map overlays. We used root cause analysis to explore drivers of performance loss.

Results: A total of 112 smears were processed (18 Coimbatore, 17 Pondicherry, 77 Madurai). Overall AI accuracy declined to 58.9% (66/112), a 20-point drop from the 79% development benchmark. Site specific AI accuracy was 38.9% (Coimbatore), 52.9% (Pondicherry), and 64.9% (Madurai), with operator level accuracy ranging from 51% to 68%. Only 2.7% (3/111) of human reads changed after viewing AI output, all downgrading fungus-positive slides to “indeterminate.” False predictions were linked to modifiable factors: loss of focus during manual scanning, air bubble artifacts from sub optimal coverslip placement, imaging delays exceeding 24h, improper shading correction and color calibration, tracking errors in MvSlide, use of 10x instead of 20x objectives, and domain shift due to training the AI model exclusively on high quality scans.

Conclusions: Deploying a high performing KOH smear AI model in routine practice resulted in substantial accuracy loss driven primarily by slide preparation and imaging workflow variability. Standardized protocols for coverslip placement, microscope calibration, and real time quality control are prerequisites for safe scaling. Current work will retrain site specific models incorporating lower quality real world images to improve robustness and close the implementation gap in ocular microbiology.