What does this research mean for the field?

The M3-Net deep learning architecture, which integrates multi-modality imaging and multi-task learning, achieves state-of-the-art accuracy and significantly reduces contouring time for organ-at-risk auto-segmentation in head and neck radiotherapy. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This study aims to develop a robust auto-segmentation framework for accurately identifying organs-at-risk in head and neck radiotherapy.

May 19, 2026Open Access

Multi-modality Multi-mask and Multi-task Auto-segmentation Network (M3-Net) for Organ-at-Risk in Head and Neck Radiotherapy

Key Points

This study aims to develop a robust auto-segmentation framework for accurately identifying organs-at-risk in head and neck radiotherapy.
Developed M3-Net, a triple-interlocked deep learning architecture.
Utilized cross-modality fusion and attention-guided feature recalibration using CT and MRI data.
Trained on 200 retrospective cases with additional 10 prospective cases for validation.
Reduced inference time by 63.6% (54823s vs. 19815s, p<0.001).
Achieved a Dice similarity coefficient greater than 0.88 for 60% of organs.
M3-Net met clinical speed and precision requirements in independent test cases.

Abstract

Purpose: Accurate segmentation of head and neck organs-at-risk (OARs) remains a critical challenge in radiotherapy planning, where current single-modality approaches often fail to address the inherent complexity of soft tissue differentiation and interpatient anatomical variations.This study aims to develop a clinically robust autosegmentation framework that synergistically integrates multimodal imaging features while optimizing computational efficiency.Methods: We present M3-Net, a triple-interlocked deep learning architecture featuring:(1) Cross-modality fusion modules with attention-guided feature recalibration between CT density maps and MRI soft-tissue contrast; (2) A hierarchical multi-mask generator producing organ-specific, regional, and global masks through parallel encoding pathways; (3) A dual-task learning mechanism combining segmentation with deformable image registration to establish voxel-level modality correspondence.The model was trained on 200 retrospective cases (160/20/20 split) with expert-reviewed contours from a tertiary cancer center, supplemented by 10 prospective cases for clinical validation.Results: M3-Net demonstrated significant improvements across three key dimensions: Efficiency: Reduced inference time by 63.6% (54823s vs. 19815s, p<0.001) through dynamic mask prioritization.These strategies improved the performance of M3-Net.Sixty percent of the organs achieved a Dice similarity coefficient (DSC) greater than 0.88.M3-Net performed best in 93.3% of all organs.It achieved the best average surface distance (ASD) for all organs.For independent test cases, the speed and precision can meet clinical requirements.Conclusion: M3-Net establishes new state-of-the-art performance for head and neck OAR segmentation, by simultaneously addressing accuracy-efficiency tradeoffs and modality discordance.The clinically validated workflow reduces contouring time by 75% while maintaining dosimetrically significant precision, enabling rapid adoption in adaptive radiotherapy protocols.

Bookmark

View Full Paper

Cite This Study

Ni et al. (Fri,) studied this question.

synapsesocial.com/papers/6a0bfda5166b51b53d378fc6 https://doi.org/https://doi.org/10.1016/j.adro.2026.102084

Bookmark

View Full Paper