Automatic Carpal Bone Segmentation (CORO / LT / SL)

Carpal bone instability occurs when the ligaments holding the scaphoid, lunate, and triquetrum together (the scapholunate (SL) and lunotriquetral (LT) ligaments) are torn, allowing these bones to shift out of alignment during wrist motion. Quantifying this motion requires tracking the position and shape of each carpal bone frame-by-frame throughout the movement, a task too time-consuming to perform manually across an entire dynamic series (1000 frames in our case).

For each plane, 20 representative frames were first selected using a k-means clustering algorithm and manually segmented to serve as training data. The bounding box shown here is propagated bidirectionally between each pair of consecutive manually segmented frames, tracking the expected position of each bone through the intermediate, non-manually-segmented frames. Within this moving bounding box, a plane-specific 2D U-Net then automatically segments the scaphoid, lunate, and triquetrum in the coronal (CORO), lunotriquetral (LT), and scapholunate (SL) planes. Together, the bounding box propagation and automatic segmentation are what enable dense, frame-by-frame inter-bone gap measurements across the full motion cycle.