Room: Karl Dean Ballroom C
Purpose: We present a machine learning-based method to further refine atlas-based segmentations using a Random Forest (RF) classifier.
Methods: An RF classifier was trained on voxels within a 2-D band Â±5mm from the boundaries of parotid glands delineated by an expert on CT scans. The classifier was trained with two different labels for voxels outside (class 0) and within (class 1) the Region of Interest (ROI), inside the band. For each voxel, training features included intensity differences from the ROI median, local first-order statistics, gradient (magnitude and direction), and position relative to the ROI centroid. A second RF-based classifier was built to identify noisy slices with dental artifacts, using Haralick texture features. The RF classifier for the ROI was applied iteratively, beginning with a band Â±1.5mm from the boundary of the initial contour on each slice, and subsequently by constructing bands around the refined contours from the previous iteration, until fewer than 2% of the candidate voxels across slices were assigned a different label. The initial multi-atlas contour was retained without alteration on slices with dental artifacts, as identified using the RF noise classifier.
Results: To validate the proposed approach, two RF classifiers (left and right parotids) were trained from CT scans of 44 head and neck cancer patients and tested on a hold-out of 10 CT scans, beginning with the results of 3 atlas-based segmentation methods (Method-I, Method-II, and Method-III) for 2 parotid structures. Our method resulted in statistically significant improvement in segmentation accuracy (p=1.47x10-7 using Dice coefficient, p=4.63x10-9 using 95th percentile Hausdorff distance) per the Wilcoxon signed rank test.
Conclusion: The addition of local RF classifiers significantly improved multi-atlas based parotid gland segmentation. We expect to incorporate this method into a clinical workflow.
Funding Support, Disclosures, and Conflict of Interest: This research was partially funded by NIH grant 1R01CA198121 and NIH/NCI Cancer Center Support grant P30 CA008748.