Generalizability Study of a Fluence Map Prediction Network

Abstract

A deep learning-based fluence map prediction network (FMPN) was developed for predicting fluence maps for given desired dose distributions. The FMPN was trained with head and neck (HN) VMAT plans only. Theoretically, the FMPN learned an inverse planning optimization procedure, so it is a general method working for other types of plans. This work is to investigate the FMPN's generalizability in various clinical scenarios apart from the training data. Materials/Methods: The FMPN which maps projections of 3D dose distribution to fluence maps was trained only with clinical HN full-arc VMAT plans. The baseline performance of the FMPN was evaluated on 102 HN full-arc VMAT plans independent of training dataset. To evaluate how well it generalizes to other clinical scenarios, we designed three tests, each of which has a feature different from training scenario. In Test A, we test the FMPN on a different treatment site, prostate. In Test B, we test the network on different delivery modality with three partial-arc VMAT plans and one IMRT plan. The first partial-arc plan is for a simple pelvis case with small PTV, while the other two are derived by blocking some angles of a plan for a complicated HN case with large concave PTV. The arc coverage for three plans are 120-360, 0-180, and 0-90/180-270 degrees. The IMRT plan is re-optimized from the HN case using eight equi-spaced angles. In Test C, we test the network on different degrees of modulation (DOM), in which four plans are generated by re-optimizing the HN case using four levels of penalty on DOM. Since there's a trade-off between plan DOM and plan quality, these four plans also distribute in a wide range of plan quality. The FMPN's baseline performance and generalizability were quantified by comparing the dose of FMPN-predicted fluence maps to the ground truth dose (the desired dose, input of FMPN) using 3D Gamma passing rate. Results: All results are listed in Table 1. As shown in Test A, FMPN achieved a prediction accuracy in prostate site as high as in training site (HN). In test B, performance is excellent on partial-arc VMAT and degrades on IMRT, which is reasonable because IMRT is too far away from training data (full-arc VMAT) from data distribution point of view. In Test C, FMPN can predict fluence map for plans with various DOM/ quality and accuracy doesn't vary with DOM. Conclusion:The FMPN trained on HN full-arc VMAT plans achieves high accuracy on various clinical scenarios without retraining the model. Our results demonstrate that the FMPN is well generalizable to other treatment site, delivery modality and various DOM, showing great potential for clinical use.