Click here to


Are you sure ?

Yes, do it No, cancel

A Convolutional Neural Network with ACGAN Augmented Data for Treatment Response Prediction Using Longitudinal Diffusion-Weighted MRI

Y Gao1*, V Ghodrati1 , A Kalbasi1 , J Fu1 , D Ruan1 , M Cao1 , C Wang1 , J Lewis1 , D Low1 , M Steinberg1 , P Hu1 , Y Yang1 , (1) David Geffen School of Medicine at UCLA, Los Angeles, CA,


(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: To predict radiotherapy treatment effect scores for localized soft tissue sarcoma patient using longitudinal diffusion MRI.

Methods: Thirty soft tissue sarcoma patients treated with five-fraction hypofractionated radiation therapy(RT) were enrolled. Diffusion-weighted MRI (DW-MRI) was acquired three times throughout the RT course using a 0.35T MR-guided radiotherapy machine. Treatment effect scores(TE) ranging from 0-100% were obtained from the post-RT surgery as a surrogate of patient treatment response. Patients were divided into three classes: C1:TE≤20%, C2:20%
Results: The average training accuracy and validation accuracies were 94.3% and 90.1% respectively, indicating that the generated samples were a good representation of the original patient data. Among the five rounds of testing, slice by slice prediction accuracy was 92.3%, 79.6%, 88.2%, 92.9%, and 85.7% respectively. The overall accuracy was 87.7%. If the majority score of all slices was assigned for the patient, the model accurately predicted the treatment effect for all patients except for one that had two out of four slices classified incorrectly.

Conclusion: This study demonstrates the potential to use deep learning to predict response from longitudinal DW-MRI. Accuracies of 87.7% and 95% were achieved on independent test sets for slice-based and patient-based prediction, respectively.

Funding Support, Disclosures, and Conflict of Interest: The authors acknowledge research support from ViewRay Inc.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email