The challenge
Neurodegenerative diseases can cause subtle changes in tissue structure. The question was whether classical MRI texture measures, computed per brain region, can separate three diagnostic groups in a small, imbalanced dataset.
The solution
T1 and FLAIR images and lesion maps were brought onto the same grid, and a cortical and subcortical brain atlas was fitted to each image. For each region and slice, GLCM texture measures, wavelet features and local binary patterns were computed, and a Random Forest classified the participants.
How it was built
The project was carried out in a group; I contributed across data processing, modelling and evaluation. The data was MRI from 125 participants (41 healthy, 61 with Alzheimer’s and 23 with Lewy body dementia). Evaluation used patient-grouped 3-fold cross-validation, keeping all slices and regions from one person in the same fold, with class weighting for the imbalance. Slice orientations were also compared.
Result
The best configuration reached a mean accuracy of about 0.66 across folds, but a macro-F1 of only around 0.54: the model recognised Alzheimer’s well, while Lewy body dementia was rarely detected.
Context
A small, imbalanced classification experiment without a separate test set; the slice orientation was chosen on the same folds, so the figures are optimistic. The method uses handcrafted texture features, not deep learning, and is not a diagnostic tool.