MPF-Net: A multi-scale feature learning network enhanced by prior knowledge integration for medical image segmentation
Yucheng Wang, Si Liang, Linyan Xue, Kexuan Zhou, Wenlong Fan, Xin Yu Cui, Minghui Wang, Shuang Liu, Kun Yang, Shilong Chang · Alexandria Engineering Journal · 2025
s Precise delineation of medical images plays a crucial role in advancing automated diagnostic systems and therapeutic strategy development. Despite the advancements in traditional CNN-based segmentation methods, they encounter significant hurdles, primarily the limited capability in capturing long-range dependencies due to the inherent localization of convolution operations, and the reduced segmentation accuracy resulting from uniform down sampling when extracting diverse scale features. We present MPF-Net, an integrated architecture that systematically addresses these limitations through designed to boost the efficiency and robustness of medical image segmentation. MPF-Net is composed of three integral components: (1) a prior information branch that employs super pixels to filter out redundant information and integrate key edge details as prior knowledge; (2) parallel convolution blocks that effectively extract diverse scale features and local context from medical images, accommodating their varying shapes and sizes; and (3) a channel-wise cross-fusion attention block, which is based on Transformer architecture, designed to capture long-range dependencies and diminish semantic gaps. Extensive experiments on three medical image segmentation datasets demonstrate MPF-Net's effectiveness, with DSC scores of 81.84 %, 91.10 %, and 90.73 % achieved on MoNuSeg, GLaS, and ISIC2018 datasets. Further evaluation on the external PH2 dataset yields a DSC of 78.73 %. MPF-Net delivers high-precision and robust generalization capabilities for segmenting complex medical images.