Session Item

The model-based approach is clinically used in the Netherlands to refer a patient to conventional radiotherapy (XT) or to a proton therapy (PT) treatment. In this context, the benefit of using deep learning (DL) models to predict XT and PT plans are substantial. The plan generation becomes automatic and the planners’ various expertise has no longer an impact. In this study, we develop an automated model-based approach by combining the use of such DL dose prediction models for intensity-modulated radiation therapy (IMRT) and pencil beam scanning (PBS) treatment with a normal tissue complication probability (NTCP) model. The accuracy of this clinical decision tool will be evaluated.

Two U-Net architectures with dense connections were trained: one for XT plans and one for PT plans. They were trained on a database of 40 esophageal cancer patients using cross-validation and a circulating test set to predict a dose distribution for each patient. The U-Net models take the CT scan, the contours of OARS and of the target volume as inputs. We also investigated an alternative approach where a first estimation of the dose distribution (D_0) is passed along with the other inputs in the DL models. D_0 is generated on the RayStation software by running a few optimization steps on a template of OARs and target objective functions. No manual fine-tuning is needed.

The DL models were chained with a NTCP model for postoperative pulmonary complications. The model parameters are the age, the histology, the BMI and the mean lung dose (MLD). The accuracy of our DL models for patient referral will be evaluated by using ∆NTCP between XT and PT plans. The protocol from the Dutch Society for Radiotherapy and Oncology recommends the use of a model-based approach with a ∆NTCP threshold value ≥ 10% to refer a patient to PT.

Our DL models succeed in predicting the dose distributions. Dice coefficients between the prediction and the manually generated plan (groundtruth) range from [0.844,0.946] for the XT model without prior-knowledge (standard model) and [0.902,0.98] for XT with prior-knowledge. For PT, the standard model reaches Dice values in the range [0.917,0.956] while it is in the range [0.945,0.981] for the prior-knowledge model. Figure 1 shows the clinical decision that is made while chaining our DL networks with the selected NTCP model. All patients are correctly classified with both approaches.

Figure 1: Clinical decision diagram representing the preferred modality for each patient. ∆NTCP predicted is the difference between the NTCP values computed from the XT and PT plans. ∆NTCP groundtruth is the difference between the NTCP values computed from the XT and PT manually generated plans.

In this work, we designed an automated clinical decision tool which consists in chaining DL dose prediction models with a NTCP model for pulmonary complications and we demonstrated the accuracy of this tool in predicting the preferred treatment modality for esophageal cancer patients.