Rib Segmentation and Sequence Labeling via Biaxial Slicing and 3D Reconstruction

Abstract

The process of diagnosing rib lesions involves radiologists interpreting 2D CT images produced by a CT scanner. To identify the location of the lesion and make an accurate diagnosis, hundreds of 2D CT images are meticulously reviewed and ribs are classified. This study proposes Transverse and Frontal Rib Segmentation (TFRS) to address the issues of labor-intensive process, and performs Sequential labeling based on it. TFRS trains 2D images composed of Transverse and Frontal planes from the chest CT volume in the U-Net model. The combination of segmentation masks produced by the model complements spatial information from different planes, reconstructing a 3D rib volume. The performance of TFRS is evaluated using Dice, Recall, and Precision metrics, showing Dice of 90.29, Recall of 89.74, and Precision of 90.72. Sequential labeling is evaluated using the Successful Labeling rate, determining whether the 12 pairs of ribs within the chest volume have been accurately labeled in sequence. The performance of Sequential labeling based on TFRS demonstrated that out of 460 test sets, 448 were correctly labeled.