Michal Franczel
Supervisor(s): Lukáš Hudec
Slovak Technical University
Abstract: Histological examination is a crucial component of breast cancer diagnostics. Analysis of whole-slide images (WSI) is a time-consuming process due to their hierarchical nature and size, resulting both in slower diagnostics and a lack of annotations. Recent advances in vision transformers have demonstrated potential within the field of computer vision. However, their properties with hierarchical gigapixel images, where contextual information is crucial, remains underexplored. In this paper, we propose a solution employing semi-supervised learning based on a self-supervised pre-training and supervised fine-tuning paradigm, utilizing these advancements. Our approach modifies vision transformer encoders within the segmentation network to incorporate contextual information from lower magnification levels through late feature fusion. The multi-scale model variant outperforms its single-scale counterpart, improving the dice score by 6.2%. Furthermore, we examine the properties of features learned by masked image modeling (MIM) and establish that vision transformers trained with MIM can effectively learn morphological phenotypes from unlabeled histopathological images, thereby validating its use as a pretraining technique in this domain.
Keywords: Computer Vision, Image Processing
Full text: Year: 2023
