Enhancing Skin Lesion Segmentation via Martingale Feature Fusion and Adaptive Deep Semantic Modeling

Skin lesion segmentation from medical images is crucial for early disease detection, yet it faces challenges due to complex textures and blurred boundaries. This paper introduces the Martingale Feature Fusion Network (MFFNet), a novel segmentation framework that leverages martingale-basedstatistical texture modeling to enhance feature expressiveness. MFFNet integrates a texture martingale module for robust texture representation, a cross-attention fusion module for multi-modal feature interaction, and a deep semantic fusion module for dynamic feature response adjustment.Experiments on ISIC2016, ISIC2017, and ISIC2018 datasets demonstrate that MFFNet outperforms existing hybrid architectures, particularly in challenging scenarios, achieving state-of-the-art performance with mean Intersection over Union (mIoU) scores of 0.8601, 0.7853, and 0.8177, respectively.These results validate the effectiveness of martingale-based texture modeling in improving segmentation accuracy. The source code is available at https://github.com/lyao519/MFFseg.