Click here to


Are you sure ?

Yes, do it No, cancel

A Deep Triplet Network with Cox Regression for Survival Prediction From Magnetic Resonance Imaging (MRI) in Patients with Soft Tissue Sarcoma

P Thammasorn1, S Schaub2, D Hippe3, M Spraker4, J Peeken5, 6, L Wootton2, P Kinahan2,3, S Combs5,6, W Chaovalitwongse1,3, M Nyflot2,3*, (1) University of Arkansas, Department of Industrial Engineering, Fayetteville AR (2) University of Washington, Department of Radiation Oncology, Seattle, WA (3) University of Washington, Department of Radiology, Seattle, WA (4) Washington University in St. Louis, Department of Radiation Oncology, Saint Louis, MO (5) Technical University of Munich, Department of Radiation Oncology, Munich, Germany (6) Helmholtz Zentrum Muenchen, Neuherberg, Germany


(Tuesday, 7/14/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: Soft tissue sarcomas (STS) consist of over 70 distinct histologies with variable behaviors, presenting challenges for patient management. There is considerable interest in development of data-driven biomarkers for precision oncology of STS patients. Here, we develop and evaluate a deep learning triplet network with Cox regression to estimate STS patient survival from magnetic resonance images (MRI).

Methods: 272 T1-weighted MRI of STS patients were retrospectively analyzed at two institutions in the US and Europe. Primary tumors were segmented by oncologists. The modified deep survival triplet network, a convolutional neural network with integrated triplet representation learning for survival analysis, was used to estimate hazard ratios for patient overall survival. The impact of the number of image features extracted from the network was evaluated (ranging from 4 –128 features). Triplet network performance was evaluated as the concordance-index (c) on the test dataset and compared to two state-of-the-art neural networks for survival analysis (DeepConvSurv and DeepSurv) and a statistical model (Cox regression of 45 texture features).

Results: The triplet network was trained on 180 patients (US cohort), tuned on 20 patients (US cohort) and tested on 72 patients (EU cohort). Performance was greatest for the triplet network with 32 learned image features (c=0.79). Overall, the triplet network had the best performance (c=0.63—0.79 depending on feature dimension). Performance of DeepConvSurv ranged from c=0.56—0.74 while DeepSurv had c=0.41—0.68, which was somewhat better than the statistical model (c=0.60) and inferior to the triplet model. Further, the triplet network significantly stratified survival of high-risk and low-risk patients on Kaplan-Meier analysis (p=0.009).

Conclusion: The triplet network yielded the best performance in predicting outcomes in an externally-validated patient cohort. Going forward, radiomic models validated in prospective trials may provide new strategies for precision treatment of sarcoma patients, such as selecting high-risk patients for treatment intensification.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email