SFCSA-YOLO: Scale Fusion and Channel Standard-deviation Attention Based YOLO for Skin Lesion Segmentation

Download PDF
Open Webpage

Shuanglong Xie, Naiyao Liang, Zuyuan Yang, Kan Xie

Published: 2024, Last Modified: 05 Nov 2025IoTAAI 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Segmenting skin lesions is essential for diagnosing and treating skin diseases. Recently, YOLOv8 can segment skin lesions, but it faces significant limitation. The main issue is its shortcoming in utilizing the complementary information between large and small-scale features. In this paper, we propose SFCSA-YOLO, which is an improved model based on YOLOv8. It incorporates a multi-scale 3D fusion module and a multi-scale alignment fusion module to utilize the potential complementary information across large and small feature scales. Furthermore, the Channel Standard-deviation Attention Mechanism (CSAM) is utilized to incorporate the standard deviation to better capture feature variability and difference, enhancing the recognition ability of multi-scale fused feature maps. Experimental results on the ISIC2018 and ISIC2016+PH2 datasets show that SFCSA-YOLO achieves state-of-the-art performance, with a 3% improvement in the index of Jaccard Similarity Index (JSI).