DSIA-Net: A dual-stream interactive aggregation network with shape prior for small target image segmentation with application to knee cartilage delineation

Accurate segmentation of knee joint cartilage in magnetic resonance imaging (MRI) remains challenging due to artifacts, noise interference, and the thin-layer structure of cartilage (average thickness: 1.5–3.5 mm) with indistinct anatomical boundaries, making precise delineation difficult even for experienced radiologists. While various computer-aided diagnostic (CAD) systems have been developed, most fail to adequately capture the diverse cartilage morphologies in MRI, resulting in suboptimal segmentation performance and limited clinical interpretability. To address these limitations, this study proposes DSIA-Net, a novel end-to-end deep learning network featuring a dual-stream interactive aggregation architecture and a shape prior module (SPM) for precise knee cartilage segmentation. The dual-stream encoder employs complementary branches—one ensuring feature stability and the other providing adaptive flexibility—with a dynamic aggregation module (ADIA) intelligently fusing these heterogeneous features. The SPM incorporates multi-scale shape prior knowledge at the decoder stage to enhance cartilage shape perception and model interpretability. Additionally, a cross-branch interaction augmentation module (CBIA) facilitates effective multi-scale feature interaction, further improving thin-layer cartilage recognition. Experimental results on datasets with varying cartilage morphologies demonstrate that the proposed method achieves competitive segmentation accuracy while maintaining clinical interpretability, showing strong potential for practical clinical applications. Supplementary materials are provided in the Appendix and available online at: https://drive.google.com/file/d/1mKtLTsljNFbgKABbY1ryAPJdTlFQo1ED/view?usp=sharing.