Bell's Palsy Severity Detection

Bell's Palsy severity is graded by clinicians from facial asymmetry. this classifies that grading into four levels from images, and reaches 98.79% accuracy on the test split.

the architecture is ResNet50 with the last twenty layers unfrozen for transfer learning, trained with Adam, cross-entropy loss, early stopping and learning-rate scheduling.

the dataset was the real constraint. 1,000+ usable images augmented out of 14,000+ candidates with rotation, zoom, flip and shear, partly to generalise, mostly to correct a class imbalance that would otherwise have made the whole thing meaningless.

judged on confusion matrices, precision, recall, F1 and ROC-AUC, because with imbalanced medical classes the headline accuracy is the least informative number available.