Dual-domain MIM based contrastive learning for CAD of developmental dysplasia of the hip with ultrasound images

Existing B-mode ultrasound (BUS) based computer-aided diagnosis (CAD) for developmental dysplasia of the hip (DDH) is mainly developed based on the Graf's method by segmenting crucial anatomical structures. However, their diagnosis performance heavily depends on the accuracy of segmentation algorithms. To this end, the pioneering CAD models based on non-Graf's method have shown their feasibility and effectiveness for DDH diagnosis. However, the deep neural network based models generally suffer from the issue of small sample size. In this work, a novel hybrid multi-task self-supervised learning algorithm, named Dual-Domain Masked Image Modeling (MIM) based Contrastive Learning (DDMCL), is proposed to improve the diagnosis performance of CAD model for DDH with limited BUS training samples. The DDMCL performs both the spatial domain MIM and frequency domain MIM to learn more comprehensively intrinsic representations from BUS images, and then integrates the dual-domain features into the contrastive learning (CL) framework. Moreover, a new hybrid loss function, i.e. the energy and mutual information based loss (named EM-Loss), is developed to jointly optimize the dual-domain network branches in CL, which can effectively distinguish the dual-domain features with large discrepancy. The experimental results on two DDH BUS datasets indicate that the proposed DDMCL outperforms all the compared algorithms, suggesting its effectiveness.