An interactive deep-learning workflow for head and neck gross tumour volume segmentation.

Background and purpose: Deep learning (DL)-based auto-segmentation of head and neck cancer (HNC) gross tumour volumes remains challenging due to anatomical complexity and limited accuracy. We propose an interactive DL (iDL) workflow that integrates clinician input to enhance segmentation performance and clinical usability. The iDL approach was evaluated using simulations on two datasets and an observer study. Materials and methods: Two iDL approaches were developed and integrated into a workflow: (1) for primary tumour (GTVt) segmentation, clinicians marked the tumour centre and delineated three orthogonal slices to enable patient-specific fine-tuning of a pre-trained 3D UNet; (2) for lymph nodes (GTVn), clinician-provided clicks identified involved lymph nodes used as attention maps in a separate UNet. Methods were evaluated using independent simulations on an internal dataset (n = 204) and the HECKTOR 2022 dataset (n = 524), using aggregated dice-similarity-coefficient (DSCagg). An additional observer study with three radiation oncologists assessed usability and efficiency, using normalized added path length (APL) and the System Usability Scale (SUS). Results: The iDL workflow achieved high segmentation accuracy, with a DSCagg of 0.84 (internal) and 0.88 (HECKTOR) for GTVt, and 0.83 and 0.85 for GTVn. GTVn required minimal correction (mean APL: 4 % vs. 6 % vs. 11 %); two observers made limited corrections to GTVt (mean APL: 11 % vs. 6 % vs. 39 %). Mean segmentation time was 12 min per case. SUS scores ranged from 87.5 to 100, indicating high usability. Conclusion: The iDL workflow achieved high accuracy and usability with limited correction time, offering a practical and efficient solution for HNC segmentation in radiotherapy.