Multi-scale Main-auxiliary MRI Sequences Fusion Based U-Net for Focal Lesion Segmentation

• Xibin Jia,
• Chuanxu Yang,
• Yifan Yang,
• Chen Qian,
• Luo Wang,
• Zhenghan Yang,
• Hui Xu,
• Xianjun Han,
• Hao Ren,
• Xinru Wu,
• Boyang Ma,
• Dawei Yang,

Vol. 19, No. 3, March 31, 2025
10.3837/tiis.2025.03.012, Download Paper (Free):

• Medical and Biological Imaging
• Multiple MRI Sequences Fusion
• Focal Lesion Segmentation
• Dual Attention

Abstract

Multi-parameter Magnetic Resonance Imaging (MRI) plays a significant role in clinical diagnosis by providing complementary information from auxiliary sequences to the main sequence. As a result, recent studies have been focusing on using multiple MRI sequences for focal lesion segmentation. Nevertheless, the challenge arises because of the lack of manual labels on some MRI sequences due to blurred lesion boundaries. So, we propose a multi-scale main-auxiliary sequences fusion based U-Net that utilizes the complementary information from both the unlabeled auxiliary sequences and the labeled main sequence for focal lesion segmentation. On one hand, to effectively augment the feature discriminability of the main sequence with the aid of auxiliary sequences, this paper proposes the multi-scale main-auxiliary sequences adaptive fusion based encoding path, where a main-auxiliary sequences adaptive fusion (MASAF) and its enhanced version Auxiliary Sequences Contribution Aware Fusion (ASCAF) are put forward to achieve main-auxiliary fusion at each scale. On the other hand, the multi-scale boundary feature enhancement-based decoding path is devised to decode the multi-scale fused features to acquire the final segmentation result. To improve discrimination capability of obscure lesion areas from the healthy tissues, a Dual Attention Module is performed at each scale to enhance the features via a channel attention and a spatial attention module. Comprehensive experiments on a self-collected clinical focal liver lesion dataset and the public dataset BraTS 2015 demonstrate that our proposed method outperforms comparative methods from both quantitative and visualized analysis.

Statistics

Cite this article