MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Physician Specific Post-Op Prostate Cancer CTV Delineation Using Deep Learning

A Balagopal*, M Lin , R Hannan , D Nguyen , S Jiang , Medical Artificial Intelligence and Automation Laboratory, Department of Radiation Oncology, University of Texas Southwestern, Dallas, Texas, USA

Presentations

(Sunday, 7/14/2019) 5:00 PM - 6:00 PM

Room: 303

Purpose: Physicians use intraoperative reports and histopathological findings to delineate the post-operative prostate cancer CTV as there is no gray scale variation observed over a large part of volume on CT images due to the microscopic disease. Although there are some guidelines for segmenting CTV, essentially there is no ground truth and the CTV contour varies significantly among physicians. The purpose of this work is to design a physician-specific deeplearning network for CTV and organ segmentation on post-operative male pelvic CT images.

Methods: The architecture consists of a 2D localization network followed by a 3D segmentation network for volumetric segmentation of CTV, bladder and rectum. The localization network is a standard 2D UNet with 3 channels, one for each organ. We use a 3D UNet-like architecture for volumetric segmentation. The network was trained on 367 post-operative cases which were contoured by 5 physicians. In order to make the segmentation physician-specific, a 3D classification network was used for classifying the styles of different physicians. The network was able to classify 4 physician styles with ~80% accuracy. Transfer learning was used for finetuning the trained network on all the physicians to the style of each of the 4 physicians using data contoured by a specific physician.

Results: The network trained on all physician data achieves Dice similarity coefficient (DSC) 84±3% for the CTV with prediction time of less than 3 mins per patient. The physician specific networks achieved an average accuracy of 3% more than the general network.

Conclusion: In the case of post-operative CTV which does not have a difference in terms of contrast or gray level over a large part of its volume on CT, the network exhibits good performance in terms of DSC compared to existing ATLAS based methods and shows improved performance when adapted to specific physician’s style.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email