Impact of annotation imperfections and auto-curation for deep learning-based organ-at-risk segmentation.

Abstract

Background and purpose: Segmentation imperfections (noise) in radiotherapy organ-at-risk segmentation naturally arise from specialist experience and image quality. Using clinical contours can result in sub-optimal convolutional neural network (CNN) training and performance, but manual curation is costly. We address the impact of simulated and clinical segmentation noise on CNN parotid gland (PG) segmentation performance and provide proof-of-concept for an easily implemented auto-curation countermeasure.

Methods and materials: The impact of segmentation imperfections was investigated by simulating noise in clean, high-quality segmentations. Curation efficacy was tested by removing lowest-scoring Dice similarity coefficient (DSC) cases early during CNN training, both in simulated (5-fold) and clinical (10-fold) settings, using our full radiotherapy clinical cohort (RTCC; N = 1750 individual PGs). Statistical significance was assessed using Bonferroni-corrected Wilcoxon signed-rank tests. Curation efficacies were evaluated using DSC and mean surface distance (MSD) on in-distribution and out-of-distribution data and visual inspection.

Results: The curation step correctly removed median(range) 98(90-100)% of corrupted segmentations and restored the majority (1.2 %/1.3 %) of DSC lost from training with 30 % corrupted segmentations. This effect was masked when using typical (non-curated) validation data. In RTCC, 20 % curation showed improved model generalizability which significantly improved out-of-distribution DSC and MSD (p < 1.0e-12, p < 1.0e-6). Improved consistency was observed in particularly the medial and anterior lobes.

Conclusions: Up to 30% case removal, the curation benefit outweighed the training variance lost through curation. Considering the notable ease of implementation, high sensitivity in simulations and performance gains already at lower curation fractions, as a conservative middle ground, we recommend 15% curation of training cases when training CNNs using clinical PG contours.