University of Texas Southwestern Medical Center at Dallas Dallas, TX
Y. Zhang, M. Zarenia, C. Sarosiek, A. Amjad, R. Conlin, B. A. Erickson, W. A. Hall, and E. S. Paulson; Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI
Purpose/Objective(s): Deep-learning based auto-segmentation (DLAS) techniques have been gradually introduced into the clinic to enhance the efficiency of the contouring process. The Automatic Contour Quality Assurance (ACQA) process plays a crucial role in ensuring safety and efficiency, particularly for online adaptive radiotherapy (OART). Previously reported population based ACQA models are less successful when applied to daily abdominal MRIs due to significant image and organ anatomy variations. This study aims to integrate patient-specific prior information (previously acquired MRI, corresponding contours, and quality labels) to intentionally overfit ACQA models (IO-ACQA) for individual patients. Materials/
Methods: The small and large bowel contours of a total of 54 abdomen MRIs from a 1.5T MRI-Linac were utilized. Among these, 8 patients contributed two image sets, with one image included in training and the other one saved for testing of each patient-specific model. Each image has manually created ground truth contours and a set of DLAS contours. The whole process includes the following steps: 1) Image pre-processing (i.e., intensity normalization, cropping, and masking); 2) Geometric and image quality augmentation to enhance variation; 3) Contour slice quality pre-labeling into accurate and inaccurate categories using previously developed contour quality classification models; 4) Feeding all the data except for the one testing set into a convolutional neural network (CNN) for training; and 5) Model testing and evaluation using classification accuracy and the rate of misclassifying inaccurate as accurate slices. Results: After augmentation, over 10,000 slices were generated for each organ and utilized for model training. The mean accuracy ranged from 0.90 ± 0.03 ([0.88-0.97]) for small bowel to 0.85 ± 0.07 ([0.72-0.94]) for large bowel contours. The misclassification rates were within the range of [1.8-11.1] for small bowel and [2.5-12.1] for large bowel contours. Model training was conducted on a GPU server (with three Nvidia K40) and took two hours with 100 epochs. The time required for one testing case was less than 20s with an i7-6700 CPU-PC. Conclusion: The IO-ACQA tool can quickly identify accurate and inaccurate contour slices with a high accuracy from auto-segmented small and large bowel contours on MRIs. It can be seamlessly integrated into a pipeline alongside our on-going automatic contour correction methods, significantly expediting the contouring process for MRgOART.
