Efficient proton-photon patient selection via dose and NTCP prediction for head and neck cancer patients

Compared to photon therapy (XT), proton therapy (PT) can often reduce normal tissue toxicity for head and neck (HN) cancer patients, despite being a limited resource. On the other hand, clinical decision-making process to select between PT and XT (e.g., treatment planning and then plan evaluation for comparing normal tissue complication probabilities (NTCP) between XT and PT) is time-consuming and resource demanding.

This study aims to develop and validate the feasibility of an artificial intelligence (AI)-based automated method for efficient patient selection between PT and XT.

A heterogeneous cohort of 104 bilateral HN patients with auto-planned PT and XT plans was analyzed, covering diverse tumor subsites and prescription dose levels. To ensure accurate dose and NTCP prediction, a joint-modality prediction framework was developed, incorporating a 3D attention-gated U-net with a multi-constrained loss function. A stratified 10-fold cross-validation strategy was employed to evaluate and compare model performance. The NTCP differences between XT and PT for grade II/III xerostomia/dysphagia exceeding certain thresholds are used to select patients for PT according to the Landelijk Indicatie Protocol Protonentherapie (versie 2.2) (LIPPv2.2).

AI-assisted patient selection process took about 10.1 s per patient. Our method achieved an accuracy of 85.58% and a weighted accuracy of 81.11% in patient selection. For dysphagia grades ≥ 2 and ≥ 3, the predicted results exhibited consistent selection with the ground truth in 86.54% and 89.42% of cases, respectively. Compared to previous models, the average ΔNTCP prediction error (ΔNTCP ground truth-ΔNTCP predicted, mean ± SD) of the proposed method was 1.47 ± 1.80%, statistically lower than U-net (1.67 ± 2.20%) and hierarchically densely connected U-net (2.34 ± 3.25%). Moreover, the joint-modality prediction of PT and XT dose distributions using Attention U-net achieved comparable performance to separate single-modality predictions.

This study highlights the potential of a novel AI-assisted framework with joint-modality prediction to enhance efficiency and precision for proton-photon patient selection in the heterogeneous dataset, demonstrating the generalizability and robustness of the proposed approach.