Funded by: Afeyan Foundation
PI: Varduhi Yeghiazaryan
Researchers: Yeva Gabrielyan, Vahe Petrosyan, Sona Bezirganyan
The project considers 3 different challenges of state-of-the-art medical image segmentation: (1) limited and/or incomplete training data, e.g. scribbles or tiny datasets, (2) inconsistent training data, e.g. multi-rater masks with significant disagreement, (3) lack of datasets for understudied medical targets. The project aims to make progress towards solving these challenges by proposing and experimentally verifying novel approaches that combine techniques from traditional computer vision, self- and semi-supervised learning, evidential deep learning, etc.
Weakly supervised learning with scribble annotations uses sparse user-drawn strokes to indicate segmentation labels on a small subset of pixels. This annotation reduces the cost of dense pixel-wise labeling, but suffers inherently from noisy and incomplete supervision. Recent scribble-based approaches in medical image segmentation address this limitation using pseudo-label-based training; however, the quality of the pseudo-labels remains a key performance limit. We propose PLESS, a generic pseudo-label enhancement strategy which improves reliability and spatial consistency. It builds on a hierarchical partitioning of the image into a hierarchy of spatially coherent regions. PLESS propagates scribble information to refine pseudo-labels within semantically coherent regions. The framework is model-agnostic and easily integrates into existing pseudo-label methods. Experiments on two public cardiac MRI datasets (ACDC and MSCMRseg) across four scribble-supervised algorithms show consistent improvements in segmentation accuracy.
Overview of the proposed PLESS framework, showing pseudo-label construction, waterfall-based enhancement, and integration into pseudo-label loss. The red dotted arrows indicate the original pseudo-label DSC loss formulation used prior to PLESS, which is replaced by the green-arrow loss terms introduced by the proposed enhancement strategy.
Analysis of the pituitary gland on MRI is clinically important for assessing endocrine disorders, but remains time-consuming and inconsistent due to the gland’s small size and variable anatomy. We present a fully automated deep learning pipeline for pituitary gland segmentation and height measurement on T1-weighted MRI. A curated dataset of 56 T1-weighted brain MRI scans was assembled from a subset of the publicly available IXI dataset, and all scans were manually annotated by expert radiologists to provide ground-truth 3D segmentations for training and evaluation. Using nnU-Net, we compared models trained on full volumes and deterministically cropped regions of interest (ROI) to quantify accuracy-efficiency trade-offs, and further evaluated test-time augmentation (TTA) to improve robustness. Our best model achieved radiologist-level segmentation accuracy (DSC = 0.845) with surface distances within expert variability, and ROI-based inference was substantially faster than full-volume processing. A principal-component-analysis–based estimator of gland height reached sub-millimeter agreement with radiologists (MAE = 0.45 mm), outperforming alternative geometric methods by accounting for gland tilt.
Deep learning models have shown strong performance in medical image segmentation. But their integration into clinical practice has been slow, largely due to the lack of reliable uncertainty estimates. In medical imaging, uncertainty arises not only from the input, but also from variability in annotations. Most existing multi-rater segmentation approaches focus on modeling label disagreement through probabilistic outputs, without providing explicit uncertainty estimates. We propose Trusted Multi-Rater Segmentation (TMS), a novel algorithm that integrates evidential deep learning into multi-rater segmentation. TMS treats network outputs associated with each annotator as subjective opinions, represented as parameters of the Dirichlet distribution, and combines them using averaging belief fusion from subjective logic. Unlike prior methods, TMS produces both probabilistic segmentations and explicit, interpretable uncertainty estimates. We show state-of-the-art segmentation of the optic disc and cup using the RIGA dataset, as well as lung nodules using the LIDC dataset. We explicitly evaluate the uncertainty estimates, showing that TMS exhibits strong uncertainty-awareness.