Work

A lightweight dual path Kolmogorov-Arnold convolution network for medical optical image segmentation

Year: 2025

Type: article

Abstract: Medical optical image segmentation (MOIS) significantly aids physicians in diagnosing and monitoring the disease by providing valuable information about the necessary anatomical regions. However, MOIS frequently has high computational costs because of its complex morphology, rich edge information, and high resolution. To this end, we propose a lightweight dual-path architecture U-Net called KAU-Net. The newly proposed network achieves reuse and re-exploration of historical information through information interaction between image feature paths and historical information paths, enabling deep layers to learn comprehensive features. To satisfy the lightweight requirements of MOIS, Kolmogorov-Arnold convolution (KAConv) is used to replace traditional convolution, thereby obtaining more information while greatly reducing computational costs. Based on KAConv, we further propose a dual-path information fusion (DPIF) module to improve segmentation accuracy. This module establishes information fusion mechanisms across paths, network layers, and stages, enabling KAU-Net to fully utilize and re-explore historical information. To further enhance the information extracted by the dual-path encoder, a multi-scale information fusion (MSIF) module is designed, effectively connecting the encoder and decoder. Compared to the state-of-the-art approach, KAU-Net obtains IoU metrics of 85.32 %, 85.02 %, 82.87 %, and 63.15 % on the four tasks and has Dice of 93.84 %, 91.25 %, 90.12 %, and 85.10 %, respectively. Moreover, only 15.15 M parameters, 1.72 G FLOPs and 504 M memory are required. The proposed method has the advantages of lightweight, versatility, and high segmentation accuracy. Our study provides a reference solution for the lightweighting of MOIS models.

Source: Neurocomputing

DOI: 10.1016/j.neucom.2025.131776

Authors: Jun Yuan, Lijun Zhou, Meirong He, Changyu Luo, Junran Zhang

Institutions: Sichuan University, Northwest Minzu University

Language: English

Topic: Medical Image Segmentation Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funders: National Natural Science Foundation of China, Sichuan University, Science and Technology Program of Gansu Province, Chengdu Science and Technology Program

Awards: G1248550418, G2409245411, G4496090550, G6307433967