Rib segmentation in chest X-ray images with oblique projection data

Background: Because computer-aided diagnoses can identify features that are not easily visible, their application in X-ray images of the chest is continuously increasing.However, most chest X-ray computer-aided diagnosis studies have been conducted using frontal projection,-such as posterior-anterior or anterior-posterior-radiographs, because other projections that might show otherwise obscured characteristic features are not widely accessible.Accordingly, we investigated the rib semantic segmentation in chest X-ray images projected frontally and obliquely.Materials and Methods: Chest X-rays were captured in frontal and oblique projections, and three UNet-based models were employed for efficient segmentation.Chest X-ray images may exhibit overlapping tissues, which can negatively affect the results.To overcome this issue, rib enhancement preprocessing was carried out, using a publicly available deep-learning rib suppression model.Results: The U2Net model was found to be the most effective model when working with oblique data, achieving a Dice coefficient of 0.89.Significantly outperforming the UNet and attention UNet, the U2Net P-values of 0.03, respectively. Conclusion:The difference in performance between oblique and frontal projections was hypothesized to be less pronounced than that between other projections.Despite employing a limited set of data with the frontal projection, we expected a significant performance improvement with the oblique projection.This hypothesis was confirmed, revealing a significant difference with P<0.05.Further research on different projections may serve as a foundation for offering more diversified information to radiologists and researchers.