Dosimetric Predictors for Quality of Life After Prostate Stereotactic Body Radiation Therapy via Deep Learning Network

Abstract

Potential correlation between the dosimetric quantities and the clinical outcome may exist, and remains a challenging task in clinical outcome assessment. To account for complex linear/non-linear relationships, we developed a Deep Neural Network (DNN) model that transforms the patient geometry and treatment parameters into corresponding quality metrics. The purpose of this work is to predict clinical outcome, i.e., patient-reported quality-of-life (QOL) in urinary and bowel domains, via novel DNN for prostate stereotactic body radiotherapy (SBRT). The DNN model was developed to predict clinical QOL scores assessed from the Expanded Prostate Cancer Index Composite (EPIC-26) at the time between consultation and initial treatment (baseline) and at the completion of treatment and every 3 months up to 30 months. Patient geometry and dose-volume histogram parameters were extracted from Eclipse treatment plans for a group of 86 patients. The evaluated dosimetric parameters included: prostate/bladder/rectal volumes; V100 (volume receiving 100% prescription dose), V50, V80, V90; maximum/mean doses to prostate/bladder/rectum. The data was annotated with each patient’s subjective quality metrics documented at 3-month intervals. The annotated data was then split between training and test cohorts as 75% and 25%, respectively. This corresponded to a group of 60 patients used for training purposes, while the remaining 26 patients were used for evaluation purposes. The training data was fed to the DNN, where a series of weights and biases applied to the plan parameters were optimized through stochastic gradient descent (SGD), using an adaptive sub-gradient method with dynamic learning rates. The number of neurons in the hidden layers was optimized for the best result, ultimately settling at 200. This learning process minimized a loss function between network-predicted clinical outcomes and the patient-reported ground-truth, by iterating over randomly sampled sets of training data. The DNN successfully predicts QOL scores. Training on the dosimetric metrics: V50, V80 and V90 values of the rectum, 90.8% accuracy were observed in predicting the QOL in rectal domain. The same metrics in urinary domain yielded 87.3% accuracy for urinary functions. Large prostate volume predicted QOL decrements in urinary functions at 92% accuracy. When QOL scores at 1-year follow-up time were included in network training along with rectum values (V50, V80, and V90), the accuracy of the neural network-predicted QOL values at 3, 6, and 9 months follow-up were 94%, 92.5%, and 89.2%, respectively. The deep learning network was adapted to predict the patient-reported QOL scores versus dosimetric parameters for prostate SBRT. Our results showed that neural network was able to predict the QOL scores for each of the patients with +/- 5 points, which supports the model’s usage in radiotherapy.