LMBF-Net: A Lightweight Multipath Bidirectional Focal Attention Network for Multifeatures Segmentation

TL;DR

This work tackles multifeature retinal segmentation by introducing LMBF-Net, a lightweight multipath bidirectional focal attention network built on a U-Net–style encoder–decoder with bidirectional skip connections. It integrates a Multipath Residual Block for multiscale feature extraction and a Focal Modulation Attention Block to refine encoder features before decoding, and employs a patch-based training strategy to address data scarcity and class imbalance. Evaluations across DRIVE, STARE, CHASE, HRF, and IDRiD show that LMBF-Net achieves state-of-the-art or competitive performance with only about 0.191M learnable parameters and fast inference times, demonstrating strong generalisability for both vascular and DR lesion segmentation. The approach offers practical potential for real-time, multifeature retinal analysis in clinical workflows, enabling robust detection of vessels, microaneurysms, hemorrhages, and exudates across diverse datasets.

Abstract

Retinal diseases can cause irreversible vision loss in both eyes if not diagnosed and treated early. Since retinal diseases are so complicated, retinal imaging is likely to show two or more abnormalities. Current deep learning techniques for segmenting retinal images with many labels and attributes have poor detection accuracy and generalisability. This paper presents a multipath convolutional neural network for multifeature segmentation. The proposed network is lightweight and spatially sensitive to information. A patch-based implementation is used to extract local image features, and focal modulation attention blocks are incorporated between the encoder and the decoder for improved segmentation. Filter optimisation is used to prevent filter overlaps and speed up model convergence. A combination of convolution operations and group convolution operations is used to reduce computational costs. This is the first robust and generalisable network capable of segmenting multiple features of fundus images (including retinal vessels, microaneurysms, optic discs, haemorrhages, hard exudates, and soft exudates). The results of our experimental evaluation on more than ten publicly available datasets with multiple features show that the proposed network outperforms recent networks despite having a small number of learnable parameters.

Figures (5)

• Figure 1: The proposed LMBF-Net architecture. (a) Bio-Net architecture xiang2020bio. (b) LMBF-Net architecture.
• Figure 2: The proposed LMBF-Net architecture. (a) Multipath Residual Block (MRB). (b) Focal Modulation Attention Block (FMAB).
• Figure 3: Illustration of the ablation analysis conducted on the DRIVE dataset for retinal vascular segmentation. (a) RGB image input. (b) Corresponding ground truth. (c) Baseline network Bio-Net output. (d) Bio-Net++ output. (e) LMBF-Net output.
• Figure 4: Segmentation results of our LMBF-Net on representative test images from the IDRiD dataset. Top to bottom row: input images, ground truth, LMBF-Net results, and zoomed views of selected regions. Left to right column: hard exudates, soft exudates, microaneurysms, haemorrhages, and optic disc. False positive pixels are coloured red, while blue pixels show false negatives.
• Figure 5: Visual comparison of sample segmentation results of the proposed LMBF-Net and recent networks on the IDRiD dataset. (a) Input RGB images. (b) Corresponding ground truths. (c) Results of LMBF-Net. (d) Results of CARNet guo2022carnet. (e) Results of Bio-Net xiang2020bio. (f) Results of foo2020multi.