LVS-Net: A Lightweight Vessels Segmentation Network for Retinal Image Analysis

TL;DR

This work introduces LVS-Net, a lightweight encoder–decoder designed for retinal vessel segmentation to enable fast, cost-efficient disease screening. The model employs multi-scale encoder blocks and a bottleneck that combines Focal Modulation Attention with Spatial Feature Refinement Blocks, while the decoder enhances multi-scale representation via SFRB-enabled skip connections. Evaluations on DRIVE, CHASE-DB, and STARE show competitive Dice scores (around the mid-80s to high-80s) with a compact footprint (~0.71M parameters and ~29.60 GFLOPs) and high AUC values (0.993–0.998). The approach also extends to retinal arteries–veins segmentation on the RITE dataset, maintaining strong performance in a multi-class setting. An ablation study confirms the contribution of multiscale blocks, SFRB, and FMAM, highlighting the synergistic benefit of FMAM and SFRB in achieving robust vessel delineation with a lightweight model.

Abstract

The analysis of retinal images for the diagnosis of various diseases is one of the emerging areas of research. Recently, the research direction has been inclined towards investigating several changes in retinal blood vessels in subjects with many neurological disorders, including dementia. This research focuses on detecting diseases early by improving the performance of models for segmentation of retinal vessels with fewer parameters, which reduces computational costs and supports faster processing. This paper presents a novel lightweight encoder-decoder model that segments retinal vessels to improve the efficiency of disease detection. It incorporates multi-scale convolutional blocks in the encoder to accurately identify vessels of various sizes and thicknesses. The bottleneck of the model integrates the Focal Modulation Attention and Spatial Feature Refinement Blocks to refine and enhance essential features for efficient segmentation. The decoder upsamples features and integrates them with the corresponding feature in the encoder using skip connections and the spatial feature refinement block at every upsampling stage to enhance feature representation at various scales. The estimated computation complexity of our proposed model is around 29.60 GFLOP with 0.71 million parameters and 2.74 MB of memory size, and it is evaluated using public datasets, that is, DRIVE, CHASE\_DB, and STARE. It outperforms existing models with dice scores of 86.44\%, 84.22\%, and 87.88\%, respectively.

Figures (7)

• Figure 1: Architecture of the proposed LVS-Net: begins with convolutional operations followed by a decoding path using transposed convolutions. Key elements include Focal Modulation Attention and Spatial Feature Refinement block in various stages of model for precise segmentation.
• Figure 2: Schematics of the proposed blocks: (a) Focal modulation attention module with context aggregation, (b) Spatial feature refinement block.
• Figure 3: Segmentation outcomes of selected test images from the DRIVE dataset. Arranged in a left-to-right sequence are the input images, specifically images 1, 2, and 19 from the dataset. Following the input images are the ground truth and the outputs of the G-Net Light, MultiResNet, SegNet, U-Net++, and LVS-Net models. True positives are shown with green color, red pixels represent false positive detections, whereas blue pixels indicate false negative detections.
• Figure 4: Segmentation outcomes of selected test images from the STARE dataset. The images displayed in the following order are: input images (specifically images 2, 3, and 5 from the dataset). Following the input images are the ground truth and the outputs of the G-Net Light, MultiResNet, SegNet, U-Net++, and LVS-Net models. True positives are shown with green color, red pixels represent false positive detections, whereas blue pixels indicate false negative detections.
• Figure 5: Segmentation outcomes of selected test images from the CHASE_DB dataset CHASEDataset. Arranged horizontally, the images displayed are as follows: the input images (specifically, images 1, 2, and 3 from the dataset). Following the input images are the ground truth and the outputs of the G-Net Light, SegNet, and LVS-Net models. True positives are shown with green color, red pixels represent false positive detections, whereas blue pixels indicate false negative detections.