Uncertainty-aware automatic segmentation with interactive refinement of head and neck squamous cell carcinoma and pathological lymph nodes

Ziping Chu, Sonit Singh, Lois Holloway, Arcot Sowmya

Keywords: Head and Neck tumour segmentation, Human-in-the-loop medical image segmentation, Head and neck squamous cell carcinoma, Uncertainty-aware interactive segmentation

TL;DR: This study proposes HN-UISeg, an uncertainty-aware automatic segmentation and interactive refinement framework for head and neck squamous cell carcinoma and pathological lymph nodes.

Abstract: Precise voxel-level annotation of head and neck tumours is crucial for radiotherapy planning but remains challenging due to the small size and irregular shape of tumours and low tumour-tissue contrast in the head and neck region. This study proposes HN-UISeg, an uncertainty-aware automatic segmentation and interactive refinement framework for head and neck squamous cell carcinoma and pathological lymph nodes. The framework was developed using the public HECKTOR and H&N1 datasets and evaluated on a local radiotherapy cohort with paired pre-treatment whole-body PET/CT and planning CT scans. HN-UISeg supports both PET/CT and CT-only input volumes, enabling tumour segmentation even when PET is unavailable. A clinically-oriented pre-processing strategy was adopted, which achieved comparable segmentation performance while maintaining explainability and plug-and-play capability. Uncertainty estimates derived from a combination of Monte Carlo dropout and test-time augmentation were used to generate voxel-wise uncertainty and probability maps, which guide an interactive refinement stage driven by simulated user clicks. Across primary tumours and nodal metastases, HN-UISeg delivers competitive automatic performance and further gains under interactive refinement, supporting more reliable and efficient contouring in head and neck radiotherapy planning. Code will be released on completion of this project.

Primary Subject Area: Segmentation

Secondary Subject Area: Uncertainty Estimation

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 201