A novel family model for dose prediction in esophageal cancer VMAT planning

The tumor distribution in esophageal cancer exhibits high heterogeneity, making the design of corresponding volumetric modulated arc therapy (VMAT) plans challenging and time-consuming for medical physicists.

This study proposes a new family model driven by multi-medical physics prior knowledge to provide clinically acceptable VMAT dose references for esophageal cancer.

This study used a training set of 505 esophageal cancer patients and 40 cases of esophageal cancer data from three centers as the testing set. Another 43 cases were used for ablation experiments and prospective evaluation. The anatomical and dosimetric prior knowledge are incorporated as constraints to guide the model in individualized predictions of VMAT dose distributions for esophageal cancer. The new family model comprises three generations of networks. First, a basic model analyzes the deep features within the dose prior knowledge, saving the parameters obtained from feature learning. These parameters, combined with anatomical prior knowledge, are then passed to the second-generation model, which serves as a pedagogical model to establish mapping relationships between anatomical and dosimetric prior knowledge. Finally, the dosimetric related parameters are removed, and a third-generation learning model independently explores potential effective features within the anatomical prior knowledge to generate the predicted VMAT dose distribution.

The absolute dose differences between the predicted and ground truth spatial dose distributions within the planning target volume (PTV) were quantified using D98%, D2%, and Dmean. Compared to state-of-the-art (SOTA) models, the new model demonstrated lower values of 49.01 cGy ± 17.93 cGy, 13.94 cGy ± 4.62 cGy, and 9.84 cGy ± 5.51 cGy for D98%, D2%, and Dmean, respectively. In terms of dosimetric evaluation for organs at risk (OARs), it also performed better than other SOTA models. Prospective evaluations revealed that the new model enables medical physicists to save at least 35.3% of their planning time compared to conventional workflows.

The novel artificial intelligence approach holds promise in providing medical physicists with valuable guidance for VMAT planning optimization.