Atlas-based bone segmentation from lower extremity MRI images

Segmentation of musculoskeletal (MS) structures from medical images is crucial for disease diagnosis, treatment planning, and computational modeling. However, current manual or semi-automatic methods are time-consuming, highlighting the need for automation. This study proposes a multi-atlas-based method to segment five bones from lower extremity magnetic resonance imaging (MRI) images. To enhance performance, we introduce two novel techniques. First, an adaptive atlas selection (AAS) method removes atlases with poor registration, improving segmentation accuracy. Second, a fast similarity computing (FSC) method based on bone probability maps reduces computational cost. The AAS technique improved sensitivity by 10.53%, Dice similarity coefficient by 5.00%, and Intersection-over-Union by 5.63% compared to existing atlas-based approaches. Moreover, FSC reduced the time for similarity computation by approximately 95%. These findings demonstrate that the proposed framework can streamline patient-specific MS model generation and facilitate broader clinical applications in MS modeling.