Research and build a labeling dataset to develop algorithm for segmenting and noting rib arcs. Specifically, after collecting from 245 scans, chest X-ray images were partitioned and labeled with 20 different bone arches. This dataset, then, is the basis for training today’s most advanced deep learning algorithms (such as U-Net, FPN, U-Net++).
Combining a standard dataset and a solid technology platform, a fully built AI model is able to accurately determine the position of the ribs from chest X-ray images. The test shows that the accuracy of the segmentation task reaches 83.4% on the Dice-score coefficient.
This result has important implications for the diagnostic process of some cardiopulmonary diseases as well as surgical intervention. Not only reducing the workload for doctors when determining the cavity of the lesion, the model also aims to be applied in solving the problem of removing ribs on X-ray images, thereby helping to read images and diagnose disease becomes easier. In particular, in the work, the skeleton labeling dataset was opened to the community to promote related research.