Program Information
Automatic Region of Interest ROI Detection for Prostate Cases Using Deep Neural Networks
T Rozario*, M Chen , S Jiang , W Lu , UT Southwestern Medical Center, Dallas, TX
Presentations
MO-RAM-GePD-J(A)-4 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: Automatic region of interest (ROI) detection in prostate cancer cases enables fast identification and accurate mapping between the critical structures and standardized label templates. This can alarm potential naming errors and facilitates automatic treatment planning. We present a novel deep learning approach that accomplishes these objectives.
Methods: A supervised learning approach using artificial neural networks (ANNs) were used to study the spatial and structural patterns of the ROIs including the tumor site and their mutual relationships. For this study, we included the PTV/GTV, prostate, bladder, Rectum, L/R femur head and body. The training data consisted of composite images generated by fusing ROI masks that were derived from DICOM RT structures of prior treated patients. One composite image was generated for each ROI per patient. To distinguish between ROIs, the pixel intensities were manipulated, such that, the ROI in consideration had pixel value 1, all the other ROIs had value 0.5 and everything else 0. Next, we down-sampled and vectorized the composite images and fed them into the ANN which generated the relevant features to capture the global relationships among the ROIs. The trained model then predicted the ROI labels to the standard ones. We implemented a two-hidden layered ANN (128, 64 nodes) using Googles Tensorflowr0.12.
Results: The multiclass ANN was trained and tested on 100 IMRT prostate cases (70%-30% training-test split) with an accuracy of 92%/100% for 30/75 epochs, respectively. The ANN was also able to identify missing -and erroneous contours of ROIs on the composite images. The accuracy reduced to 99% and 96% with simulated local and global pixel translations of 3% and 5%, respectively.
Conclusion: The results indicate potential use of deep learning in auto-planning for accurate and efficient structure sorting. Future work involves testing the model on Head&Neck cases and achieving robustness.
Contact Email: