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

Developmental dysplasia of the hip (DDH) is a critical congenital condition in infants, which is generally early diagnosed by B-mode ultrasound (BUS). Deep learning-based computer-aided diagnosis has shown effectiveness in diagnosing DDH from BUS images. However, the issue of small sample size remains a challenge. To this end, a Dual-domain Cross-constraint Masked image modeling based hybrid Symmetry Convolution-Mamba Network (DCM-SCMNet) is proposed for the BUS-based diagnosis of DDH with limited training samples. In the pretext task, dual-domain (spatial and frequency) Masked Image Modeling (MIM) is designed using a shared hybrid encoder and two decoders. A special dual-domain cross-constraint strategy is then developed for the dual-domain reconstruction of two decoders, which can benefit to learning more inherent feature representation. Furthermore, a novel SCMNet is proposed as the encoder in MIM, which integrates symmetry convolution and a novel symmetry Mamba into a block to effectively capture more inherent anatomical symmetry of hip in BUS images. In particular, the symmetry Mamba designs a bidirectional scanning mechanism tailored for the properties of bone structure symmetry. It then can effectively learn feature representation and still maintain computational efficiency. Experimental results demonstrate that DCM-SCMNet outperforms all the compared algorithms on the DDH dataset.