ESTRO 2024 - Abstract Book

S3129

Physics - Autosegmentation

ESTRO 2024

Automated auditing of clinical usage of deep learning auto-segmentation can track changes in user behaviour over time. Results will be used to develop a QA tool to highlight when the editing of a deep learning contour is outside the normal range.

Keywords: Deep learning segmentation, clinical audit

References:

[1] Nikolov, Stanislav, et al. "Deep learning to achieve clinically applicable segmentation of head and neck anatomy for radiotherapy." arXiv preprint arXiv:1809.04430 (2018).

2715

Proffered Paper

Empowering robust tumour segmentation in multi-centre OPC cohorts: DL tumour probability maps

Alessia De Biase 1,2 , Nanna M Sijtsema 1 , Roel J H M Steenbakkers 1 , Johannes A Langendijk 1 , Lisanne V van Dijk 1 , Peter M A van Ooijen 1,2 1 University Medical Centre Groningen, Radiation Oncology, Groningen, Netherlands. 2 University Medical Centre Groningen, Data Science Centre in Health (DASH), Groningen, Netherlands

Purpose/Objective:

The clinical adoption of auto segmentation methods in radiotherapy treatment planning faces significant hurdles, with the lack of transparency in deep learning (DL) segmentation techniques and the notable inter-observer variability in manual target delineation standing out as primary impediments. A novel DL-based method was developed to assist radiation oncologists in primary gross tumour volume (GTVp) segmentation in oropharyngeal cancer (OPC) patients. The tool provides the model’s predictions and visually includes a measure of the model's confidence in its prediction. In other words, instead of generating a single fixed contour, the DL-based approach produces a predicted tumour probability for each PET-CT voxel, hence resulting in a 3D probability map. The current study had two primary objectives: 1) to test the performance of the DL-based method for auto segmentation of both the GTVp as well as the pathologic lymph nodes (GTVln) based on PET/CT images of OPC patients in an internal and external test set; 2) to demonstrate the advantage of predicting not only one fixed tumour contour but a complete 3D tumour probability map, making the method more robust for differences in patient and image characteristics across different cohorts.

Material/Methods:

The previously published DL segmentation network (De Biase et al., 2023) was shown to perform well in GTVp (single label) auto segmentation, hence it was extended, in this study, to perform auto segmentation of both GTVp and GTVln