A Clinically Guided Approach for Training Deep Neural Networks for Myopic Maculopathy Classification

Abstract

Pathologic myopia (PM) is a sight-threatening disease characterised by abnormal ocular changes due to excessive axial elongation in myopes. One important clinical manifestation of PM is myopic maculopathy (MM), which is categorised into 5 ordinal classes based on the established META-PM classification framework. This paper details a robust deep learning approach to automatically classifying MM from colour fundus photographs as part of the recently held Myopic Maculopathy Analysis Challenge (MMAC). A ResNet-18 model pretrained on ImageNet-1K was trained for the task. Pertinent MM lesions (patchy or macular atrophy) were manually segmented in images from the MMAC dataset and another publicly available dataset (PALM) to create a collection of lesion masks based on which an additional 250 images with severe MM were synthesised to mitigate class imbalance in the original training set. The image synthesis pipeline was guided by clinical domain knowledge: (1) synthesised macular atrophy tended to be circular with a regressed fibrovascular membrane near its centre, while patchy atrophy was more irregular and varied more greatly in size; (2) synthesised images were created using images with diffuse or patchy atrophy as background; and (3) synthesised images included examples that were not easily classifiable (e.g. creating patchy lesions that were in close proximity to the fovea). This, coupled with mix-up augmentation and ensemble prediction via test-time augmentation, enabled the model to rank first in the validation phase and fifth in the test phase. The source code is freely available at https://github.com/fyii200/MyopicMaculopathyClassification.

F. Yii is supported by the Medical Research Council [grant number MR/N013166/1]. The funder had no role in the design and conduct of this work nor the decision to submit this manuscript for publication.