Exhibit Hall | Forum 4
Purpose: Survival prediction for the patients undergoing stereotactic body radiotherapy (SBRT) for liver metastases remains an unsolved task. In this study, we present a multi-modal neural network (MDNN) for survival prediction in patients undergoing liver SBRT.
Methods: An IRB-approved retrospective review was performed at our institution for patients with metastatic liver lesions who underwent liver SBRT between 2012 and 2021. Outcome of interest was 18 months overall survival (18mos-OS). We have designed a MDNN composed of two pathways: (1) A 1D pathway that takes a vector of clinically relevant features including age, percentage of normal liver volume, distance to the hilum, and EQD2, and (2) A U-Net-derived pathway that takes as input a volumetric CT image centered around the planning target volume. After a series of fully connected layers, the 1D pathway concatenates with the U-Net-derived pathway. After concatenation, dense layers are followed by a Softmax layer where outcome prediction is made. The model was trained, validated and tested on 71 patients for whom 18mos-OS was available. Implementation of the model was done in Keras with Tensorflow backend. Accuracy defined as the ratio of correct predictions to the sample number, and the area under the curve (AUC) of receiving operator characteristics were calculated as performance metrics.
Results: The median age of the patients cohort used was 66 and the 18mos-OS was 50%. On the 15 external validation samples, the MDNN model performed an AUC of 0.81 and an accuracy of 0.74. Offline training time of the model was about 20 minutes and prediction time was about 15 ms per patient.
Conclusion: Through the deep learning-based integration of radiological, dosimetric and clinical data, a promising workflow for outcome prediction was obtained. The workflow can reinforce precision medicine in radiotherapy via adoption of outcome-directed strategies of treatment intensification or de-escalation.