Title: Prototype Based Segmentation of Bone Tissue Microscopy Images Using Self-Supervised Vision Transformers and Feature Space Similarity
Abstract:
Segmentation of microscopy images constitutes a fundamental task in biomedical research and clinical analysis. However, many segmentation methods rely on large annotated datasets. As the creation of labeled datasets tend to be highly time consuming and difficult to scale, there exists a need for finding alternative segmentation methods that can use unlabeled data directly.
This thesis investigates whether pretrained self-supervised Vision Transformers can be used for prototype based segmentation of bone tissue microscopy images. A segmentation framework based on pretrained DINOv2 backbones was developed, in which positive and negative reference points are used to construct prototype embeddings that guide similarity based segmentation in the learned feature space. The framework was evaluated using multiple DINOv2 backbone variants, feature space analysis and prototype transfer experiments.
The results demonstrated the potential of using pretrained self-supervised Vision Transformers for microscopy image segmentation by showing that the models produce feature representations in which tissue and background regions become partially separable. Despite being trained on natural RGB images rather than microscopy data, the pretrained backbones enabled segmentation of bone structures using the proposed similarity based segmentation framework.
Supervisor: Mirja Granfors Pineda
Examiner: Giovanni Volpe
Opponent: Patrik Dennis
Place: FL71
Time: 15 June, 2026, 09:00