Click here to


Are you sure ?

Yes, do it No, cancel

Learning Clinical Expertise Using Deep 3D Networks: An Automated Clinical Target Volume (CTV) Delineation for Non-Small Cell Lung Cancer (NSCLC) Patients

Y Xie1*, K Kang2, Y Wang3, F Keane4, M Khandekar5, H Willers6, T Bortfeld7, (1) Massachusetts General Hospital, Boston, MA, (2) Bio-tree System, Inc., ,,(3) Massachusetts General Hospital, Boston, MA, (4) Massachusetts General Hospital, ,,(5) Massachusetts General Hospital, ,,(6) Massachusetts General Hospital, Boston, MA, (7) Massachusetts General Hospital, Boston, MA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: CTV delineation can be the most difficult step in the process of defining the treatment volumes; and it accounts for the high workload for most radiation oncology departments worldwide. We developed an automated CTV delineation network for locally advanced NSCLC. Different with organ-at-risk segmentation based merely on images, our system learned to delineate using both patient specific anatomic and clinical variables, as well as physicians’ clinical experience.

Methods: network takes both patients’ CT and their corresponding ITV as inputs, and outputs CTV delineation. The ITVs were contoured by physicians from the 4D CT simulation scans. From fifty-four patients with locally advanced NSCLC in the right lung treated with standard chemotherapy and radiation, forty-six datasets were randomly selected to train the network, and the other eight were used for testing. Then, we compared the predicted CTVs with physicians' contours from two aspects: 1) visually inspected the network predicted CTV for avoiding critical organs, such as heart and esophagus; and 2) quantitatively measured the missed and overlapped volume by precision and recall, respectively.

Results: CTV contours delineated by computer avoided critical organs-at-risk. Seven out of eight CTV predictions visually followed closely with the physician’s CTV contours. Only one had large discrepancies near the apex of the upper lobe, but acceptable by physicians. The performance of the trained network was also evaluated quantitatively using precision and recall. The average precision was 0.893, and recall was 0.834.

Conclusion: automated CTV delineation was utilized for right lung NSCLC patients using the deep learning method solely based on CT images with manually contoured ITV. On the limited testing samples, the model predicted CTVs reasonably well. This approach demonstrates the capability of an intelligent volumetric expansions with consideration of the patient specific anatomic and clinical variables, as well as the clinical expertise of the radiation oncologists.


3D, Image-guided Therapy, Treatment Planning


Not Applicable / None Entered.

Contact Email