# Label Quality

The study proposes HierEM, a hierarchical expectation-maximization framework that treats each site's observed prostate lesion annotation as a noisy observation of a latent clean lesion mask, alternating between inferring a voxel-wise posterior over that latent mask and training a CNN with the posterior as a soft target while estimating site- and case-level sensitivity and specificity under a logistic-normal hierarchical prior; on three-site data it reaches per-site mean DSC of 29.50%–39.69% in pooled held-out evaluation and 27.91%–32.67% in leave-one-site-out generalization, with statistically significant improvements over comparison methods (p < 0.039) and interpretable per-site label-quality estimates (sensitivity α of 31.5%–47.3% at specificity β ≈ 0.99).

The study proposes HierEM, a hierarchical expectation-maximization framework that treats each site's observed prostate lesion annotation as a noisy observation of a latent clean lesion mask, alternating between inferring a voxel-wise posterior over that latent mask and training a CNN with the posterior as a soft target while estimating site- and case-level sensitivity and specificity under a logistic-normal hierarchical prior; on three-site data it reaches per-site mean DSC of 29.50%–39.69% in pooled held-out evaluation and 27.91%–32.67% in leave-one-site-out generalization, with statistically significant improvements over comparison methods (p < 0.039) and interpretable per-site label-quality estimates (sensitivity α of 31.5%–47.3% at specificity β ≈ 0.99).

Across four abdominal CT datasets (WORD, AMOS, CT-1K, AbdomenAtlas), the authors generated pseudo-label variants with seven models (nnU-Net, MedSAM, TotalSegmentator, and four STU-Net sizes), then trained DynUNet under identical deterministic settings for in-domain training and for pretraining followed by fine-tuning; in-domain performance rose strongly and non-linearly with label quality and dataset volume partly compensated for quality, whereas fine-tuned models significantly outperformed a no-pretrain baseline in the vast majority of settings and pretraining label quality no longer clearly affected downstream results.

Across four abdominal CT datasets (WORD, AMOS, CT-1K, AbdomenAtlas), the authors generated pseudo-label variants with seven models (nnU-Net, MedSAM, TotalSegmentator, and four STU-Net sizes), then trained DynUNet under identical deterministic settings for in-domain training and for pretraining followed by fine-tuning; in-domain performance rose strongly and non-linearly with label quality and dataset volume partly compensated for quality, whereas fine-tuned models significantly outperformed a no-pretrain baseline in the vast majority of settings and pretraining label quality no longer clearly affected downstream results.