Quantifying White Matter Hyperintensities: Predicting Periventricular Fazekas Scores with Uncertainty Estimation

White matter hyperintensities (WMH) are crucial markers in brain magnetic resonance (MR) images, often quantified using the Fazekas score. This study presents a deep learning model to predict Fazekas scores in the periventricular region from T1-weighted and FLAIR images. Our model achieved a Matthew correlation coefficient of 0.68 and F1 scores of 0.69, 0.88, 0.86, and 1.0 for Fazekas scores from 0 to 3, respectively, surpassing previous methods. We introduce a novel integration of t-distributed stochastic neighbor embedding (t-SNE) with uncertainty analysis using Monte Carlo dropout, offering insights into the model's decision-making. Results show effective class distinction, with increased uncertainty at class transitions indicating ambiguity, and confident misclassified cases suggesting overlapping features or label noise. These findings highlight that simple, well-tuned models, coupled with interpretability techniques, can provide robust predictions of WMH severity.