Input feature design and its impact on the performance of deep learning models for predicting fluence maps in intensity-modulated radiation therapy

Abstract

Objective: Deep learning (DL) models for fluence map prediction (FMP) have great potential to reduce treatment planning time in intensity-modulated radiation therapy (IMRT) by avoiding the lengthy inverse optimization process. This study aims to improve the rigor of input feature design in a DL-FMP model by examining how different designs of input features influence model prediction performance. Approach: This study included 231 head-and-neck (HN) IMRT patients. Three input feature designs were investigated. The first design (D1) assumed that information of all critical structures from all beam angles should be combined to predict fluence maps. The second design (D2) assumed that local anatomical information was sufficient for predicting radiation intensity of a beamlet at a respective beam angle. The third design (D3) assumed the need for both local anatomical information and inter-beam modulation to predict radiation intensity values of the beamlets that intersect at a voxel. For each input design, we tailored the DL model accordingly. All models were trained using the same set of ground truth plans (GT plans). The plans generated by DL models (DL plans) were analyzed using key dose-volume metrics. One-way ANOVA with multiple comparisons correction (Bonferroni method) was performed (significance level=0.05). Main results: For PTV-related metrics, all DL plans had significantly higher maximum dose (p<0.001), conformity index (p<0.001), and heterogeneity index (p<0.001) compared to GT plans, with D2 being the worst performer. Meanwhile, except for cord+5mm (p<0.001), DL plans of all designs resulted in OAR dose metrics that are comparable to those of GT plans. Significance: Local anatomical information contains most of the information that DL models need to predict fluence maps for clinically acceptable OAR sparing. Input features from beam angles are needed to achieve the best PTV coverage. These results provide valuable insights for further improvement of DL-FMP models and DL models in general.