The Potential of AI-assistance and Explainability for Contouring of Tumors

Radiation Oncology has evolved rapidly in recent decades in terms of innovations in treatment equipment, volumetric imaging, information technology and increased knowledge in cancer biology. New delivery technologies and associated imaging modalities have enabled highly optimized precision radiation therapy and contributed to improvements in tumor control and cancer patient cure. Associated with these changes the growing burden of caner worldwide, along the scarcity of human resources.
The selection and contouring of target volumes and organs-at-risk (OARs) have become crucial steps in modern radiation oncology. The definitions for gross tumor volume (GTV), clinical target volume (CTV), and OARs have evolved over time and are widely accepted by the radiation oncology community. Despite guidelines from international groups, there are significant variations in contouring practices. Recently, deep learning-based AI algorithms have significantly improved auto-segmentation, reducing inter-observer variation and saving time. Previous studies have shown that AI assistance reduces contouring time and variation, even without prior training. Tumor contouring remains challenging due to imaging limitations and irregular tumor boundaries, making AI assistance potentially beneficial. However, explainable AI (XAI) is needed to help clinicians understand AI predictions.
The study aims to explore the effects of AI and XAI on tumor contouring, the impact of education on AI/XAI-assisted contouring, and the potential bias introduced by AI. Radiotherapy centers from various income levels, treating head and neck cancer, will participate, with at least 80 oncologists recruited, focusing on low- and middle-income countries. The study will use advanced head and neck cancer cases with primary and nodal target volumes, including CT and PET/CT images sourced from Aarhus, Denmark. The AI model, based on nnUNet and refined for head and neck cancer tumors, will be further fine-tuned for the study. Explainability will be achieved using probability and entropy maps for uncertainty representation.
The study will be conducted in two phases: Phase A will compare AI assistance with and without explainability to manual contouring, while Phase B will investigate the risk of bias with and without explainability. An optional Phase C will explore visualization methods for explainability. Participants will be randomized into groups for contouring tasks, with teaching sessions included. The study aims to understand the effects of AI and XAI on contouring consistency and speed, and the potential risk of bias. The hypotheses include improved consistency and speed with AI/XAI, and reduced bias with XAI.