Efficient Pyramid Channel Attention Network for Pathological Myopia Recognition

TL;DR

This work tackles PM recognition from fundus images by embedding global-local pathology distribution priors into CNN features. It introduces the EPCA module, combining pyramid pooling with multi-scale context fusion to recalibrate channel-wise representations, and builds EPCA-Net by stacking Res-EPCA blocks. The authors also collect PM-Fundus to provide a public benchmark and demonstrate that adapters within a pretraining-and-finetuning framework enable competitive PM recognition with fewer tunable parameters than traditional fine-tuning. Overall, the approach advances PM diagnosis with an efficient, distribution-aware attention mechanism and suggests a viable path for leveraging large natural-image foundation models in medical imaging under data-limited regimes.

Abstract

Pathological myopia (PM) is the leading ocular disease for impaired vision worldwide. Clinically, the characteristic of pathology distribution in PM is global-local on the fundus image, which plays a significant role in assisting clinicians in diagnosing PM. However, most existing deep neural networks focused on designing complex architectures but rarely explored the pathology distribution prior of PM. To tackle this issue, we propose an efficient pyramid channel attention (EPCA) module, which fully leverages the potential of the clinical pathology prior of PM with pyramid pooling and multi-scale context fusion. Then, we construct EPCA-Net for automatic PM recognition based on fundus images by stacking a sequence of EPCA modules. Moreover, motivated by the recent pretraining-and-finetuning paradigm, we attempt to adapt pre-trained natural image models for PM recognition by freezing them and treating the EPCA and other attention modules as adapters. In addition, we construct a PM recognition benchmark termed PM-fundus by collecting fundus images of PM from publicly available datasets. The comprehensive experiments demonstrate the superiority of our EPCA-Net over state-of-the-art methods in the PM recognition task. The results also show that our method based on the pretraining-and-finetuning paradigm achieves competitive performance through comparisons to part of previous methods based on traditional fine-tuning paradigm with fewer tunable parameters, which has the potential to leverage more natural image foundation models to address the PM recognition task in limited medical data regime.

Figures (8)

• Figure 1: Tessellated fundus information (red), parapapillary atrophy (blue) and optic disc (green) on the fundus image.
• Figure 2: The general framework of EPCA-Net for automatic PM recognition on fundus images (a). We design an efficient pyramid channel attention (EPCA) module (c) by exploiting global-local pathology distribution prior information, then combine it with residual block to form a Res-EPCA module (b).
• Figure 3: A simple implementation of pyramid pooling based on the $S_{1}$ and $S_{3}$.
• Figure 4: The implementations of Drop-EPCA module with Dropout-SCFM, Hier-EPCA with Hierarchical-SCFM, and Par-EPCA with Parallel-SCFM.
• Figure 5: TThe visual comparison between the traditional finetuning paradigm and the pretraining-and-finetuning paradigm by transferring the pre-trained natural image model of ResNet50 to tackle the PM recognition task.