Comparative Validation of Graphical Models for Learning Tumor Segmentations from Noisy Manual Annotations

Classification-based approaches for segmenting medical images commonly suffer from missing ground truth: often one has to resort to manual labelings by human experts, which may show considerable intra-rater and inter-rater variability. We experimentally evaluate several latent class and latent score models for tumor classification based on manual segmentations of different quality, using approximate variational techniques for inference. For the first time, we also study models that make use of image feature information on this specific task. Additionally, we analyze the outcome of hybrid techniques formed by combining aspects of different models. Benchmarking results on simulated MR images of brain tumors are presented: while simple baseline techniques already gave very competitive performance, significant improvements could be made by explicitly accounting for rater quality. Furthermore, we point out the transfer of these models to the task of fusing manual tumor segmentations derived from different imaging modalities on real-world data.