Click here to


Are you sure ?

Yes, do it No, cancel

Predicting Lung Tumor Shrinkage During Radiotherapy Seen in a Longitudinal MR Imaging Study Via a Deep Learning Algorithm

c wang*, Y Hu , A Rimner , N Tyagi , E Yorke , G Mageras , J Deasy , P Zhang , Memorial Sloan-Kettering Cancer Center, New York, NY


(Wednesday, 8/1/2018) 7:30 AM - 9:30 AM

Room: Karl Dean Ballroom C

Purpose: Early prediction of lung tumor shrinkage is highly desirable to facilitate decision making in adaptive radiotherapy. The purpose of this project is to develop a deep learning algorithm and predict the trajectory of lung tumor shrinkage during radiotherapy monitored via a longitudinal MRI study.

Methods: Nine lung cancer patients were monitored during radiotherapy via 6-7 weekly MRI-T2W scans under an IRB-approved protocol. Tumors on all scans were manually segmented by a radiation oncologist as the ground truth. A convolutional neural network (CNN) was designed to utilize the first three weekly scans to predict the spatial distributions of tumors in the following three weeks. After all weekly scans were rigidly registered to the first weekly scan by matching the lung contour, a patient specific region of interest (ROI) was defined as the first week tumor extended with a uniform 3D margin of 1.5cm. Around each voxel inside the ROI, three 3cmx3cm patches were generated from the first three weekly scans as the input. The prediction of the corresponding voxel on the following three weeks, background or tumor, was the output. The performance of the CNN was cross-evaluated under a leave-one-out scheme, with metrics including precision, Dice coefficient, and root mean square surface distance (RMSSD) between the predicted and actual tumors.

Results: Tumors shrank 60%±27% (mean ± standard deviation) at the end of radiotherapy across nine patients. CNN predicted tumors on week (4, 5, 6) with a precision, Dice, and RMSSD of (0.78±0.09, 0.69±0.10, 0.67±0.06), (0.81±0.06, 0.73±0.05, 0.73±0.06), and (1.8±0.8mm, 2.1±0.7mm, 2.0±0.9mm), respectively.

Conclusion: The deep learning algorithm can capture and predict the spatial shrinkage patterns of the tumor along the course of radiotherapy. It can be integrated into the clinical workflow of adaptive radiotherapy. More patient data are needed to improve and validate this algorithm.

Funding Support, Disclosures, and Conflict of Interest: MSKCC has a research agreement with Varian Medical System.


Lung, Image Guidance, MRI


Not Applicable / None Entered.

Contact Email