The central obstacle the team targeted is the small sample size that plagues medical imaging datasets. Annotated ultrasound scans of infant hips are difficult to collect in large volumes, which limits how well conventional networks can learn diagnostic feature representations. DCM-SCMNet addresses this constraint through a masked image modeling pretext task that trains the network before it ever sees a labeled diagnosis.

In the pretext task, the authors designed dual-domain masked image modeling operating in both the spatial and the frequency domain. A shared hybrid encoder processes the masked inputs, while two separate decoders handle reconstruction in each domain. A special dual-domain cross-constraint strategy then links the two decoders.

The encoder at the heart of the pretext task is a novel hybrid symmetry convolution-Mamba network, or SCMNet. It integrates symmetry convolution and a novel symmetry Mamba within a single block, an arrangement intended to effectively capture the inherent anatomical symmetry of the hip that appears in B-mode ultrasound images. The symmetry Mamba component introduces a bidirectional scanning mechanism tailored to the properties of bone structure symmetry.

According to the authors, this mechanism can effectively learn feature representation while still maintaining computational efficiency, a balance that matters for clinical deployment where processing speed and resource use are practical concerns. Experimental results reported in the paper demonstrate that DCM-SCMNet outperforms all the compared algorithms on the DDH dataset. The evaluation places the new model against existing deep learning approaches on the same B-mode ultrasound data, with the dual-domain cross-constraint pretext task and the symmetry-aware encoder credited for the gains.

The work suggests that combining masked image modeling across spatial and frequency domains with anatomy-tailored network blocks can ease the small-sample challenge in ultrasound-based computer-aided diagnosis.

Subject of Research: Deep learning architecture for ultrasound-based diagnosis of developmental dysplasia of the hip with limited training samples

Article Title: Dual-domain cross-constraint MIM based hybrid symmetry convolution-Mamba network for ultrasound diagnosis of DDH

Article References: Sun, K., Liu, J., Du, J., Shi, J., Li, J., Wang, J., & Shi, J. (2026). Dual-domain cross-constraint MIM based hybrid symmetry convolution-Mamba network for ultrasound diagnosis of DDH. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03700-2

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03700-2

Keywords: DDH, ultrasound, deep learning, masked image modeling, Mamba, computer-aided diagnosis, symmetry convolution, small sample size, medical imaging, hip dysplasia, dual-domain, cross-constraint