Room: Track 4
Purpose: This study investigates the dosimetric impact of a deep learning (DL) auto-contouring tool that encapsulates the knowledge of expert cardiovascular radiologists applied to external-radiation-therapy (RT) treatment planning.
Materials: An in-house designed and elsewhere geometrically validated DL system was trained on CT scans from cardiology (n=858) and applied to radiation oncology planning CT scans for automated heart segmentation. This retrospective dosimetry study included 2867 patients treated for breast cancer at our institution between 2016-2018. Mean heart dose (MHD) was computed for each case and used to dosimetrically validate the DL system by comparing against standard clinical variations in MHD. A random subset of 20 patients was selected and re-contoured by 20 medical professionals. For the remaining patients the DL results were compared against clinically treated heart contours.
Results: Inter-user variation in contouring led to normally distributed (p>0.05) MHD with a standard deviation of 3.3cGy. The MHD of the DL generated contours with respect to manual contouring is within one standard deviation for 68% of the cases. The maximum observed difference in MHD (dMHD) was 3.6cGy for DL and 29.5cGy for manual contouring. For the remaining 2847 patients the dMHD for the 50th, 75th and 95th percentile is found to be 2.6cGy, 5.3cGy, and 49.8cGy, respectively. All patients with a dMHD of more than 50cGy were within 3.5mm distance to the 50% isodose line. Based on a one-sided Wilcoxon signed-rank test the DL system returns significantly higher MHDs (p<0.001) as compared to manual contours.
Conclusion: This study dosimetrically validated a DL auto-contouring system. For more than 75% of the cases human input would not have changed clinical decision making (MHD below threshold of 100cGy and dMHD less than 10cGy). It demonstrates the feasibility of DL-based auto-contouring for RT planning and as a quality assurance tool that can substantially change clinical workflows