Multi-scale information residual network: Deep residual network of prostate cancer segmentation based on multi scale information guidance

• Published in: Biomedical signal processing and control Vol. 110; p. 108132
• Main Authors: Chen, Xinyi, Liu, Xiang, Yu, Yunjie, Shi, Yunyu, Wu, Yuke, Wang, Zhenglei, Wang, Shuohong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2025
• Subjects: DeepLearning; MIGF model; MSR-Net; Prostate Cancer; Segmentation; DeepLearning; Prostate Cancer; Segmentation; MIGF model; MSR-Net
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.108132

Magnetic resonance imaging (MRI) is a key tool in prostate cancer screening and diagnosis, with automatic segmentation of the cancer crucial for accurate staging and treatment. Nevertheless, the accurate segmentation of prostate cancer remains a challenging subject. In order to address this challenge, a two-stage segmentation method is employed. In the initial stage, the prostate tissue is delineated as the region of interest. Subsequently, in the second stage, the precise segmentation of prostate cancer is achieved on the aforementioned region of interest. In order to accurately segment the region of interest, we propose MSR-Net (Multi-scale information residual network), which employs an MSR-block, designed based on MLKA convolution, to extract multi-scale information, combines the group attention mechanism to enhance the multi-scale features, and uses the improved CGA feature fusion module to fuse deep and shallow features. The feature map is then upsampled using DySample. The experiments conducted on the Prostatex dataset for the segmentation of prostate cancer were based on the Dice metric. The results demonstrate an improvement of 5.2% (60.5% vs. 55.3%) in comparison to the second-best performing segmentation network (Unet). The application of the two-stage segmentation method has a considerable impact, with a 10.4% improvement (45.3% vs 55.7%) on the baseline when two-stage segmentation is employed for prostate cancer in comparison to direct segmentation of prostate cancer. Furthermore, the network was trained and tested on the prostate segmentation and lung nodule segmentation datasets, achieving the highest dice scores of 0.937 and 0.764, respectively.