Room: Stars at Night Ballroom 2-3
Purpose: Rapid and accurate segmentation is essential for online adaptive radiation therapy (OART). This study aims to develop a framework to automatically evaluate contour quality using quantitative MRI texture and shape features for MRI-guided OART.
Methods: The framework includes: (1) pre-processing images, (2) extracting texture and shape features on a slice-by-slice basis in the contours to be checked and feature changes between 4-mm inner/outer shells and a core region, (3) ranking all the features using a recursive feature elimination method and selecting top-ranked features, and (4) building and testing supervised classification models for contour validation. A variety of contours of the pancreatic-head generated on 22 sets of T1-weighted non-contrast MRIs of 11 patients was utilized to demonstrate the framework. Three sets of contours were created: (1) ground truth, (2) auto-generated contour using deformable image registration, and (3) modified contour from that in (2). The contour on a slice from set #2 or 3 was labeled as accurate if Dice similarity coefficient â‰¥ 0.85, mean distance to agreement â‰¤1.5 mm, and 95% of distance to agreement â‰¤ 5 mm, as compared to the ground truth. A total of 754 accurate and 650 inaccurate contour slices were used, where 80% of these contour slices were used for the model training and 20% for testing.
Results: A total of 132 features were extracted from each slice and ranked. The top-ranked 30 features were selected for model building where 12 different classification models were trained and compared. The best performing model yielded average sensitivity of 91% and specificity of 89% on the testing data after 5-fold cross-validation.
Conclusion: The newly-developed framework can automatically identify accurate and inaccurate contours on a slice-by-slice basis on MRI with high sensitivity and specificity, thus, can be implemented for quick contour check in MRI-guided OART.