Dose prediction for head-and-neck cancer using a fusion of lightweight 3D multi-scale feature enhancement modules

Radiotherapy is a cornerstone of head-and-neck cancer (HNC) treatment, but traditional radiation therapy planning remains time-consuming, experience-dependent, and prone to inconsistent quality. Deep learning-based dose prediction has emerged as a promising solution, yet existing models struggle with insufficient long-range spatial correlation capture and imbalanced prediction accuracy across high- and low-dose regions. Thus, there is an urgent need for a tailored framework to address these clinical challenges.

This study proposes a U-Net-based encoder-decoder model fused with lightweight 3D multi-scale feature enhancement modules for 3D HNC radiotherapy dose prediction, aiming to improve target dose precision and organs-at-risk (OARs) sparing.

A public OpenKBP dataset consisting of 340 HNC patients (200 for training, 40 for validation, 100 for testing) undergoing 6MV IMRT was utilized. Data were preprocessed (including CT truncation, normalization, multimodal integration) and augmented (random flipping, translation, rotation) to enhance generalization. The proposed model integrates an eight-layer Transformer for global feature extraction, 3D multi-scale convolutional blocks (MSCBs) for fine-grained feature capture, and an efficient multi-scale convolutional attention decoding (EMCAD) module for optimized feature fusion. A baseline DOSE-PYFER model, GAN-based models, and a cascade Transformer-based model were used for comparative analysis. Performance was evaluated using Dose score, DVH score, and gamma passing rate.

The proposed model exhibited superior accuracy compared to comparative models. It achieved a Dose score of 2.704 Gy and a DVH score of 1.611 Gy, outperforming the baseline and cascade Transformer-based model. The gamma passing rate reached 92.76%, indicating excellent spatial consistency with ground truth. No overfitting was observed, with stable training and validation loss curves. The model efficiently generates 3D dose distributions, supporting rapid clinical workflow.

The proposed model, integrating Transformer-driven global feature extraction and EMCAD-based multi-scale attention decoding, outperforms conventional models in HNC dose prediction. It effectively resolves long-range correlation capture and dose region balance issues, and can be clinically deployed to streamline radiotherapy planning, enhance plan consistency, and support timely adaptive radiotherapy.