GAMNet: a gated attention mechanism network for grading myopic traction maculopathy in OCT images

Zhou, Yan,Chen, Xiang,Lin, Shiqun,Dai, Rongping

DOI: https://doi.org/10.1007/s00371-024-03386-3

IF: 2.835

2024-05-08

The Visual Computer

Abstract:Myopic traction maculopathy (MTM) is a retinal disease caused by tractional forces on the macula, serving as a major contributor to irreversible visual impairment in highly myopic eyes. Given the clinical diversity of MTM, its classification is crucial for providing customized management and care decisions. Current methods primarily fall into two categories: Traditional approaches use machine learning for feature extraction, while deep learning methods employ data-driven training of classification models. However, due to the spatial similarity of optical coherence tomography (OCT) images, data imbalance, and coarse-grained partitioning, the MTM classification task remains challenging. Thus, we propose a gated attention mechanism network (GAMNet) to automatically grade MTM using OCT images. GAMNet adopts a ResNet-34 backbone with a selective fusion of temporal and channel attention modules to construct the basic architecture. For improved feature extraction ability, GAMNet combines a gated attention mechanism following a dual attention module. To address the robustness issues associated with data imbalance, our model utilizes a combination of data augmentation techniques and a weighted loss function. Our model was trained and evaluated by a dataset containing 26,616 OCT images collected from 2499 eyes. Comparing with six baseline models, experimental results indicate that the GAMNet successfully achieves the intended objectives, with an accuracy of 93.3%, an F1 score of 90.0%, and a recall rate of 89.7%, outperforming existing MTM classification methods.

computer science, software engineering