Show abstract - Show schedule - PDF - Reviews

Segmenting and labeling correctly the individual ribs from chest radiograph (CXR) are of significant clinical value for several diagnostic tasks. Developing automatic deep learning (DL) algorithms for this task requires annotated images of the ribs at pixel-level. However, to the best of our knowledge, there exists no such public datasets as well as benchmark protocols for performance evaluation. To solve this problem, we establish a new CXR dataset, namely VinDr-RibCXR, for automatically segmenting and labeling of individual ribs. The VinDr-RibCXR contains 245 posteroanterior CXRs with corresponding segmentation annotations for each rib provided by human experts. Furthermore, we train the state-of-the-art DL-based segmentation models on 196 images from the RibCXR and report performance of those models on an independent test set of 49 images. Our best performing DL model (i.e., Nested U-Net with EfficientNet-B0) obtains a Dice score of 0.834 (95% CI, 0.810-0.853). The sensitivity, specificity and Hausdorff distance are 0.841 (95% CI, 0.812-0.858), 0.998 (95% CI, 0.997-0.998), and 15.453 (95% CI, 13.340-17.450), respectively. These results demonstrate a high-level of performance in labeling of the individual ribs from CXRs. Our study, therefore, serves as a proof of concept and baseline performance for future research. The dataset, codes, and trained DL models will be made publicly available to encourage new advances in this research direction.
Hide abstract

Wednesday 7th July
A4-12 (short): Segmentation - 13:45 - 14:30 (UTC+2)
Hide schedule