Multicenter validation and randomized crossover reader evaluation of deep learning-assisted tri-sequence...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Radiotherapy planning

Background:

• Accurate MRI-based target delineation for hypopharyngeal squamous cell carcinoma (HPSCC) is crucial but expertise-dependent.
• A multicenter-validated deep-learning model is needed to standardize HPSCC contouring.
• Assessing AI's role in bridging contouring expertise gaps is important for clinical deployment.

Purpose of the Study:

• To develop and validate a multicenter tri-sequence deep-learning model for HPSCC MRI delineation.
• To evaluate if AI assistance reduces the expertise gap in tumor contouring.
• To explore quality-aware risk modeling for deployment support using low-overlap prediction.

Main Methods:

• A 3D nnU-Net model was trained on 530 HPSCC patients and evaluated across three institutions (727 total patients).
• Performance was assessed using Dice Similarity Coefficient (DSC), surface DSC, ASSD, and MSD.
• A randomized reader study compared manual vs. AI-assisted contouring by junior and senior radiation oncologists.

Main Results:

• The tri-sequence model achieved high accuracy with mean DSC of 0.82-0.87 across internal and external cohorts.
• AI assistance significantly improved junior oncologists' DSC (0.73 to 0.86) and reduced contouring time by 55%.
• AI improved inter-reader agreement (Fleiss' κ from 0.69 to 0.86) and the risk model achieved AUCs of 0.71-0.89.

Conclusions:

• Deep-learning-assisted tri-sequence MRI segmentation provides robust multicenter HPSCC delineation.
• AI significantly enhances contouring efficiency, consistency, and reduces the expertise gap.
• The findings support AI integration for quality-aware radiotherapy planning.