M.Sc. thesis · University of Hamburg & SDSUFlagship
Brain MRI Classification & Explainable AI
Does a high score on public brain MRI collections show what a model learned about tumors, or what it learned about the datasets?
- Type
- M.Sc. thesis
- Where
- University of Hamburg and San Diego State University
- Period
- Jun 2026 – Present
- Methods
- CNNs, Vision Transformers, Grad-CAM, HiResCAM, cross-dataset evaluation
- Stack
- Python, PyTorch
- Status
- Manuscript in preparation
Motivation
A benchmark score is a claim about a model. It needs checking.
Public brain MRI collections are a common starting point for tumor classification work, and models trained on them often report very high accuracy. A score on its own does not say what the model relied on to reach it.
In medical imaging that distinction matters. A classifier can be useful only if it responds to the anatomy in the scan. If it responds to something else that happens to differ between classes or between sources, the same score would mean much less.
Research question
The thesis asks whether high classification performance on public brain MRI collections reflects meaningful learning, or whether models exploit dataset-specific shortcuts. It looks at this from two sides: how the evaluation is set up, and where the models direct their attention.
Method
Four parts, built to be compared.
- 01
Matched models
Convolutional networks and Vision Transformers are built in PyTorch and trained and evaluated under the same protocols, so differences can be attributed to the architecture.
- 02
Evaluation protocol
The pipeline keeps training and evaluation data strictly separate and checks for leakage between them, so that a reported number describes performance on images the model has not seen.
- 03
Cross-dataset evaluation
Models are tested on collections other than the one they were trained on, which separates what transfers from what is specific to a single source.
- 04
Saliency against annotations
Grad-CAM and HiResCAM maps are compared with expert tumor annotations, to test whether the highlighted regions fall on the regions that matter.
Status
The manuscript is in preparation. Results, figures, and code will be added to this page once they are published.