Towards Improving Clinical Practice to Evaluate Rib Endpoints with Autosegmentation
Abstract
Purpose
Our current clinical practice relies on using a chest wall contour as a surrogate for rib contours, as contouring individual ribs is prohibitive due to time constraints. This process introduces uncertainty to the validity of fracture risk estimates as derived from published stereotactic body radiotherapy (SBRT) constraint tables. We aim to use an in-house trained, two-stage Dynamic Graph Convolutional Neural Network (DGCNN) for automated rib segmentation to overcome this barrier. This study validates the clinical accuracy and efficiency of automated rib segmentation deployed via built-in scripts within a treatment planning system.
Methods
A two-stage DGCNN architecture was trained on an open-source dataset based on published methodology for rib autosegmentation. The model first isolated the ribs before labeling the individual ribs. The model was trained on 320 cases and performance was tested on a validation dataset of 50 thoracic computed tomography (CT) scans held out from training. Geometric accuracy was quantified using the Dice Similarity Coefficient (DSC) and global pixel accuracy. Clinical utility was assessed by three clinicians using a 5-point Likert scale on a cohort of 10 local patients.
Results
The scripted application generated rib contours in approximately 70 seconds per scan. Quantitative analysis on the validation set (N=50) yielded a mean DSC of 0.91 and a voxel-wise accuracy of 96%. In the clinical evaluation (N=10), clinicians rated the contours as neutral, indicating that minor edits were required that would take less time than contouring from scratch, primarily near sternocostal and costovertebral joints where the model was not trained.
Conclusion
This work demonstrates that a two-stage DGCNN enables rapid, high-accuracy rib segmentation, removing a key practical barrier to individual rib-sparing planning. Replacing the chest wall surrogate with organ-specific rib contours allows more accurate dose assessment and supports improved modeling of rib fracture toxicity in thoracic SBRT.