Attention map of Multiple Disease Detection AI in Ultra-widefield Color Fundus Photographs and Agreement Rate of Retinal Tears

Abstract

Purpose : In the domain of fundus photography, two primary methodologies exist for AI-based lesion localization: one involves preparing data annotated with lesion locations, and the other involves preparing data labeled with disease names. While the latter approach simplifies dataset creation, it does not necessarily guarantee straightforward learning progress, even for distinct lesions. In this study, we utilized a dataset comprising 45,325 Ultra-Wide Field (UWF) images from four facilities using optos California equipment. The dataset, labeled with multiple disease names, was trained using ResNet. We investigated whether attention maps generated during testing, which included data containing retinal tears, accurately targeted these lesions.

Methods : Within the dataset, there were 355 Ultra-Wide Field (UWF) images containing retinal tears. We randomly allocated 10% of these images for validation purposes and another 10% for testing.

Results : For the test data, there were images of retinal tears from 14 patients, encompassing 15 eyes and totaling 35 images. Out of the 35 images analyzed, retinal tears were identified within the field of capture in 30 images. Among these, 21 images demonstrated a probability of over 40% for the presence of a retinal tear, as indicated by the heat maps.

Conclusions : In Ultra-Wide Field (UWF) imaging learning, solely based on multi-label disease annotations, attention maps overlapped with retinal tears at a confidence level of 40% or higher in 70% of the cases.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

A small retinal tear without associated detachment was observed at the 8 o'clock position of the extreme peripheral retina.

 

Responses on the attention map were also observed at the site of the retinal tear, in conjunction with the optic nerve head.