ESTRO 2024 - Abstract Book

S3157

Physics - Autosegmentation

ESTRO 2024

Systematic annotation errors are learned and reproduced by DL models more rigidly than random errors, which balance out. Curation of clinical datasets could be questioned, since only 1% model performance is lost with 10-30% noisy labels (systematic). However, since the amount, mode and severity of noise in clinical data is unknown, this cannot be relied on. In clinical data, a simple auto-curation step can lead to marginal (>0.5%) in-distribution DSC improvements and can considerably (>1.0%) improve when considering model generalizability, where this trend was consistent over multiple curation fractions for DSC and MSD. In subsampled training data, this was less strongly observed, which could be because smaller sample sets annotations are not guaranteed to contain adequate noise for curation to have an effect. Based on current experimental setting, we recommend removal of 10-20% lowest-ranking DSC performance cases in the training set. One limitation is that this could also remove correctly contoured but more difficult cases that a model would require for learning sufficient variation. Future directives involve other structures (organs-at-risk or target volumes) and curation based on a (weighted) scoring of DSC, HD and/or MSD.

Keywords: Annotation noise, auto-curation

References:

1. Raudaschl, P.F. et al . Evaluation of segmentation methods on head and neck CT: Auto-segmentation challenge 2015. Med. Phys . (2017) doi:10.1002/mp.12197

3106

Digital Poster

Race bias analysis of a deep learning-based prostate MR autocontouring model

Maram Alqarni 1,2 , Emma Jones 3,1 , Luis Ribeiro 4 , Verma Hema 5 , Sian Cooper 4 , Stephen Morris 4 , Teresa Guerrero Urbano 4 , Andrew P. King 1 1 King’s College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom. 2 Imam Abdurhamn bin Faisal University, Biomedical Engineering, Dammam, Saudi Arabia. 3 Guy’s and St Thomas’ NHS Foundation Trust, Department of Medical Physics, London, United Kingdom. 4 Guy’s and St Thomas’ NHS Foundation Trust, Department of Clinical Haematology and Oncology, London, United Kingdom. 5 Guy’s and St Thomas’ NHS Foundation Trust, Radiology, London, United Kingdom

Purpose/Objective:

Recently, there has been growing interest in addressing potential demographic bias in deep learning (DL) models when trained with data that are not representative of the diversity of patient populations, for example by race or sex. Such demographic bias has been investigated in other applications [1],[2] but not yet in radiotherapy (RT). The most closely related work was in the field of head and neck oncology [3], which investigated geographic bias. No work has yet investigated the potential for race bias in DL based prostate autocontouring. Therefore, this work aims to investigate possible race bias in a DL autocontouring model for prostate MR images in cancer treatment planning.