Enhanced diabetic macular edema detection in multicolor imaging through a multi-feature decomposition fusion attention network

Abstract

The objective of accurately diagnosing diabetic macular oedema (DME) is to minimize the likelihood of vision loss in affected individuals. The use of multicolor images (MCI), which offer a range of spectral representations of the fundus, aids in the detection of DME. Advanced deep learning algorithms have been developed to classify DME in MCI, but they often fall short in accuracy because they do not fully utilize the multifaceted characteristics of these images. This study introduces the Multi-feature Decomposition Fusion Attention Network (MDFANet), a novel approach for classifying MCI in both healthy individuals and those with DME. The MDFANet begins with Restormer blocks to extract initial features, then diverges into two pathways: one employing a Lite Transformer to capture broad, low-frequency features, and another using an Invertible Neural Network to focus on high-frequency details. Additionally, we have created a Residual hybrid attention modules module that refines feature extraction by integrating sequential Hybrid Attention Modules with residual connections in convolutions. This design leverages both extensive and localized feature information to enhance the analysis of multiple features. As a result, the MDFANet has proven to be effective in the early and accurate detection of DME, which is crucial for formulating well-timed treatment strategies and reducing the risk of vision impairment for patients.