Radiation Physics

PV QA 3 - Poster Viewing Q&A 3

TU_7_3184 - Deep Learning Based Tracking of Imaging Phenotypes to Improve Therapy Survival Prediction

Tuesday, October 23
1:00 PM - 2:30 PM
Location: Innovation Hub, Exhibit Hall 3

Deep Learning Based Tracking of Imaging Phenotypes to Improve Therapy Survival Prediction
Y. Xu1,2, A. Hosny3, T. P. Coroller4, R. Zeleznik3, R. H. Mak5, and H. Aerts5; 1Harvard Medical School, Boston, MA, 2Brigham and Women's Hospital, Boston, MA, 3Dana-Farber Cancer Institute, Boston, MA, 4Dana-Farber Cancer Institute, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, 5Dana-Farber Cancer Institute, Brigham and Women's Hospital, Harvard Medical School, Boston, MA

Purpose/Objective(s): Radiomics has demonstrated the ability to describe tumor phenotypes and clinical outcome predictions through quantitative image analysis. Previous studies have focused on the development of imaging biomarkers at a single time point. However, the tumor is a dynamic biological system which develops over time, and the phenotype may not be captured by a single time-point scan. With transfer learning and utilization of neural networks pre-trained on natural images, these networks can be applied to the medical field. We propose that the analysis of locally advanced non-small cell lung cancer (NSCLC) patients treated with radiation therapy incorporating follow-up CT scans through deep learning will lead to improved prediction of clinical survival outcomes.

Materials/Methods: CT images of 183 stage III non-surgical NSCLC patients treated with radiation therapy were analyzed. A total of 484 scans were analyzed, with an average of 2.6 (range 2-3) scans per patient, of pretreatment and follow-up scans were labeled with a seed point at the center of the tumor and were the inputs for the analysis. Transfer learning through the ResNet convolutional neural network (CNN) was used. For time series delta analysis, a gated recurrent unit was used at the end of the ResNet components to incorporate CT scans at different time points. Survival predictions were analyzed with the AUC and survival differences between high and low risk groups through the rank-sums and log-rank tests respectively. The analyses were compared to a clinical model with features of stage, gender, age, tumor grade, performance and smoking status.

Results: A CNN model based on the pretreatment scan only demonstrated low performance for predicting two year overall survival (AUC=0.53, p=0.52). Enhanced performance was observed if the CNN model combined the pretreatment scan with the first follow-up image (AUC=0.64, p<0.05). Further improvement was observed if the second follow up scan was included into the model for overall survival (AUC=0.70, p<0.01). Similar results were found for one year overall survival: AUC=0.54, 0.70, 0.73, p=0.48, p<0.05 respectively. The difference between patients through Kaplan-Meier analysis for high and low risk groups of the predictions with two follow up scans were significant (p<0.05). The clinical model did not result in a statistically significant prediction of survival.

Conclusion: This study demonstrates that transfer learning though pre-trained neural networks and incorporating patient scans at multiple time points has the potential to improve prediction of clinical survival. The addition of follow-up scans may provide additional phenotypic information to the neural network and allow for survival prediction compared to pretreatment scans alone.

Author Disclosure: Y. Xu: None. A. Hosny: None. T.P. Coroller: None. R. Zeleznik: None. R.H. Mak: Consultant; Boehringer-Ingelheim, Inc. Stock; Celgene. H. Aerts: Research Grant; National Institutes of Health. Consultant; Sphera, Genospace.

Yiwen Xu, PhD

Presentation(s):

Send Email for Yiwen Xu


Assets

TU_7_3184 - Deep Learning Based Tracking of Imaging Phenotypes to Improve Therapy Survival Prediction



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 Deep Learning Based Tracking of Imaging Phenotypes to Improve Therapy Survival Prediction