Radiation and Cancer Physics

SS 29 - Physics 9 - Imaging for Treatment Planning

209 - Prostate Cancer Staging and Radiation Treatment Planning Using Deep Learning on MRI

Tuesday, October 23
5:15 PM - 5:25 PM
Location: Room 304

Prostate Cancer Staging and Radiation Treatment Planning Using Deep Learning on MRI
B. Han, Y. Yuan, S. L. Hancock, H. P. Bagshaw, M. K. Buyyounouski, and L. Xing; Stanford University School of Medicine, Palo Alto, CA

Purpose/Objective(s): Currently, prostate cancer staging relies on biopsy, which can be inaccurate and is not risk-free for the patients. This study is to develop a novel deep-learning based automatic staging and treatment planning methods for prostate cancer with multi-parametric magnetic resonance images (MRI).

Materials/Methods: In prostate cancer, radiation treatment recommendations are reliant on accurate staging, which determines whether patients receive high dose rate (HDR) brachytherapy, external beam radiation (EBRT) or the combination. We developed a deep convolutional neural network to automatic assess the Gleason Score and stage the prostate cancer. The proposed deep learning model was composed of three branches of transfer learning, extracting and concatenating features from T2 and apparent diffusion coefficient MRI images. Subsequently, a novel loss function with a similarity constraint was designed to constrain the distribution of the features in the same category to be in a narrow angle region. In addition, with the joint supervision of softmax loss in the fine-tuning process, the deep discriminative features with both the interclass separability and the intraclass compactness were preserved and the feature margin between categories is enlarged. The accuracy, recall, and precision for the prostate cancer staging is tested. Furthermore, the areas under the receiver operating characteristic curve (AUC) was also utilized to measure the discrimination ability for prostate cancer staging.

Results: The robustness and effectiveness of the proposed deep learning model were tested on two cohorts: 112 cases from the PROSTATEx-2 Challenge; and 132 cases from our institutional review board approved patient database. 80% of the dataset were selected randomly as the training set for transfer learning and the remaining 20% images were used as the test set. We extracted features obtained in each layer and utilized the image similarity loss function to evaluate the recognition performance. The accuracy, recall and precision were all increased in the first five convolutional layers from ~0.6 to ~0.85. We compared our proposed model with several baseline methods including the transfer learning model without fine-tuning; models using a single MRI sequence; and model using traditional softmax loss function. The AUC of our method was 0.892 which is better than all other models with AUC of 0.784-0.859. Better performance was also achieved the in comparison with five other hand-craft features based and deep learning based prostate cancer diagnosis methods. The diagnosis result was then used to perform an automatic radiation treatment selection and the treatment outcome prediction was provided based on the clinical data.

Conclusion: The novel deep learning based automatic prostate cancer staging method provided accurate and valuable information for treatment outcome optimization and precision medicine.

Author Disclosure: B. Han: None. Y. Yuan: None. S.L. Hancock: None. M.K. Buyyounouski: None. L. Xing: Research Grant; Varian Medical Systems. Honoraria; Varian Medical Systems. Royalty; Varian Medical Systems, Standard Imaging Inc.

Bin Han, PhD

Disclosure:
Employment
Stanford University: Clinical Assistant Professor: Employee

Presentation(s):

Send Email for Bin Han


Assets

209 - Prostate Cancer Staging and Radiation Treatment Planning Using Deep Learning on MRI



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Prostate Cancer Staging and Radiation Treatment Planning Using Deep Learning on MRI