Breast cancer is the most common cancer worldwide, with over 2M new cases annually. In
mammography, the classification of microcalcifications (MCs) plays a key role in diagnosis,
but remains challenging due to limited resolution, low contrast, and tissue superposition.
High-resolution micro-CT enables non-destructive 3D imaging of paraffin-embedded biopsy
samples, allowing visualization of MCs and surrounding tissue in 3D. Previous work has mainly
focused on individual MC properties [1], as tissue distortion after extraction limits reliable
analysis of cluster characteristics. However, 16 µm micro-CT scans show potential for
preserving local spatial organization, enabling the study of cluster-like structures.Deep
learning (DL) methods such as convolutional neural networks (CNNs) have shown strong
performance in volumetric medical imaging but do not explicitly model spatial relationships
between nearby regions. Graph Neural Networks (GNNs) provide an alternative by
representing data as nodes and edges, enabling the modeling of spatial interactions. Prior
work has explored graph-based representations of MCs and tissue structures, showing that
incorporating spatial relationships can improve classification performance [2],[3]. In this thesis, each
paraffin block may contain multiple tissue pieces with several MC clusters. While labels are available at block
level, individual clusters can be treated as separate samples for modeling and classification.

To investigate whether explicitly modeling spatial relationships between local 3D regions
improves classification performance compared to standard CNN-based approaches.
Specifically, the study will compare:
• CNN-based classification using cropped 3D regions around MC clusters
• Graph-based classification, where 3D patches are represented as nodes and spatial
proximity defines edges

Literature review (ETOC: 2 months): Review the medical problem and current
methods for MC analysis, with a focus on CNNs and GNNs in medical imaging.
• Dataset familiarization (ETOC: 1 month): Understand the structure of the micro-CT
dataset, including paraffin blocks and MC clusters.
• CNN baseline implementation (ETOC: 1 month): Develop a 3D CNN baseline for MC
cluster classification, including preprocessing, data augmentation, and evaluation.
• Graph construction and GNN implementation (ETOC: 3 months): Extract 3D patches
from each cluster, represent them as graph nodes, define edges based on spatial
proximity, and implement a GNN model for classification.
• Evaluation and comparison (ETOC: 2 months): Compare CNN and GNN approaches
using cross-validation with grouping at block level, and analyze performance
differences.
• Thesis writing (ETOC: 1 month).

(Mandatory) qualifications:
• Following an MSc in a field related to one or more of the following: Computer Science,
Biomedical Engineering, Applied Computer Science - Digital Health.
• Strong programming skills (Python).
• Experience with image processing and DL.
• Interest/Motivation in developing state-of-the-art DL methods and conduct
experiments.