Abstract:
Rare brain disorder poses a big problem to diagnostics as it is not prevalent in the
population, has confusing symptoms, and is not well-known in clinical practice. The
classical forms of diagnostics do not support timely and accurate detection, postponing
therapy and deteriorating clinical outcomes. As a solution to this problem, this paper
builds a deep learning model to classify rare brain disorders in multiple classes by using
MRI scans. The dataset consisted of 3,759 images of four disorders, namely Fukuyama
Muscular Dystrophy (Fukuyama), Moyamoya Disease with Intraventricular Hemorrhage
(Moyamoya), Pantothenate Kinase-Associated Neurodegeneration (PKAN), and
Pachygyria with Cerebellar Hypoplasia (Pachygyria). The images were first preprocessed
and resized and then divided into training, validation and testing cases, and tested in eight
deep learning architectures, including Vision Transformer (ViT), DenseNet201, VGG19,
MobileNetV2, Xception, ResNet152, InceptionV3, and EfficientNetB7. ViT was the most
accurate, composing 99.47%, closely followed by VGG19 (98.40%) and
MobileNetV2(98.01%). These findings support the idea that transformer-based and
densely connected models are especially useful in modeling the complex spatial patterns
of rare brain disorders. Altogether, the research shows that AI-based systems have the
potential to assist clinicians in diagnosing rare neurological conditions more efficiently and
accurately and ultimately contribute to better healthcare outcomes among underserved
groups of patients.
