Šimon Ukuš
Supervisor(s): prof. Ing. Vanda Benešová, PhD.
Slovak Technical University
Abstract: Breast cancer grading via the Nottingham Grading System relies on the evaluation of nuclear pleomorphism, but automating this process is often hindered by real-world data constraints. In many practical clinical settings, detailed pixel-level annotation of nuclei is expensive and time-consuming. As a result, annotations may be sparse, imprecise, or limited to coarse geometric markers such as ellipses, leading to missing labels and weak supervision. This paper proposes a segmentation-driven framework to overcome these challenges. We first utilize a patch-based segmentation approach leveraging Cellpose-SAM, combined with an Intersection over Union (IoU) stitching strategy. This pipeline extracts precise nuclei instances from Whole Slide Images. To align these segmentation masks with imprecise expert annotations, we introduce a centroid-based label matching algorithm that mitigates localization errors. Finally, to resolve the general issue of missing labels within the regions, we introduce a semi-supervised classification framework. A Siamese neural network trained with triplet loss learns a discriminative embedding space from the matched nuclei, leveraging domain-specific histopathology pretraining. Unlabeled instances are subsequently assigned classes using nearest-centroid assignment in the learned embedding space. The proposed pipeline enables the translation of rough localization cues into fully segmented and classified nuclei instances, providing a structured dataset suitable for downstream nuclear pleomorphism assessment.
Keywords: Computer Vision, Image Processing
Full text: Year: 2026
