Biological Dose Prediction Model with Deep Neural Networks on Bulky Thoracic Cancer Patients of Partial Stereotactic Ablation Radiotherapy: For Better Therapeutic Evaluation and Risk Assessment

Abstract

Partial stereotactic ablation radiotherapy(P-SABR) could achieve encouraging local control rates for intractable bulky thoracic cancer. However, the range of ablation may increase the radiation risk of normal tissues surrounded. The purpose of this work was to develop a biological dose prediction model that considers tissue bioreaction in addition to patient anatomy to achieve more comprehensive evaluation for bulky thoracic tumor control and tissue toxicity. A newly developed deep learning architecture, Nestnet, was trained to establish the biological dose prediction model (D-B Nestnet). A database containing images and P-SABR plans of 94 bulky thoracic cancer patients was studied (74 patients for training/validation and 20 patients for testing). Patient-specific parameters of gross tumor ablation boost (GTVb) and other anatomical information, PTV and organs at risk (OAR), were extracted from the structure set. The linear quadratic (LQ)model applied to stereotactic radiotherapy and cell sensitivity for thoracic tissues in hyperfractionation radiotherapy were programmed to transfer the physical dose extracted from plans into biologically effective dose (BED). the training process was completed automatically after the validation loss continued for 100 epochs without fluctuation. We evaluated the model’s accuracy by comparing the minimum BED of GTVb and PTV, the maximum BED and the mean BED of all targets, BED-volume metrics for the predicted biological dose distribution against those of clinical delivered. We also trained a model based on the standard U-net architecture (D-B Unet) and compared the accuracy with the proposed model we mentioned before. The results demonstrate that the B-Nestnet model can predict clinically acceptable biological dose distributions. the average absolute biases of [max, mean] BED for GTVb, PTV are [2.1%, 3.3%] and [2.1%, 4.72%], respectively. Averaging across most of OARs, our proposed model is capable of predicting the errors of the max and mean BED within 6.3% and 6.1%, respectively. The standard D-B Unet performed worse, having averaged max and mean BED prediction errors within 9.9% and 9.7%, respectively. The new developed model, which predicts the biological dose instead of the physical dose, can still predict the distribution with reasonable accuracy. Using this model could provide a more intuitive prediction of tumor ablation and OARs risk. According to the prediction outcomes, clinicalists could modify the range of ablation for enhancing the BED of tumor or decreasing the risk of OARs toxicity. This novel D-B Nestnet model represents a major step forward towards P-SABR planning on bulky thoracic cancer in the real clinical practice.