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

Area of Science:

• Medical imaging and radiation oncology
• Artificial intelligence in healthcare
• Computational anatomy and segmentation

Background:

• Deep learning (DL) auto-segmentation for head and neck cancer (HNC) gross tumour volumes faces challenges in accuracy and anatomical complexity.
• Existing methods often require significant manual correction, impacting clinical workflow efficiency.

Purpose of the Study:

• To develop and evaluate an interactive DL (iDL) workflow integrating clinician input to enhance segmentation performance and usability for HNC gross tumour volumes.
• To improve the accuracy and efficiency of tumour delineation in radiotherapy planning.

Main Methods:

• Two iDL approaches were developed: one for primary tumour (GTVt) segmentation using clinician-marked centres and slices for fine-tuning a 3D UNet, and another for lymph node (GTVn) segmentation using clinician clicks as attention maps.
• The workflow was evaluated using simulations on internal and HECKTOR 2022 datasets, assessing Dice-Similarity-Coefficient (DSCagg).
• An observer study with radiation oncologists evaluated usability and efficiency using normalized added path length (APL) and System Usability Scale (SUS).

Main Results:

• The iDL workflow demonstrated high segmentation accuracy, achieving DSCagg of 0.84-0.88 for GTVt and 0.83-0.85 for GTVn across datasets.
• Minimal corrections were needed for GTVn (mean APL: 4-11%), with limited corrections for GTVt (mean APL: 6-39%).
• Mean segmentation time was 12 minutes per case, and SUS scores indicated high usability (87.5-100).

Conclusions:

• The interactive DL workflow provides a practical and efficient solution for head and neck cancer segmentation in radiotherapy.
• High accuracy, usability, and reduced correction time make the iDL approach a valuable tool for clinical application.