Abstract

Using an MRI to determine the type of hydrocephalus a person has is a crucial step in diagnosis and treatment planning. Vision transformers (ViTs) and convolutional neural networks (CNNs) have performed admirably, but they often struggle to process large amounts of input or connect disparate portions of the environment. This paper presents a novel hybrid deep-learning architecture for MRI image feature extraction, which uses EfficientNet for local feature extraction and the Vision Transformer for global dependency capture. The model is trained and tested using an MRI dataset linked to hydrocephalus. Local Interpretable Model-agnostic Explanations (LIME) are one method for dealing with the “black box” component of complex deep learning models. It promotes trust and candor among doctors by providing them with understandable visual explanations of the model’s predictions. When it comes to recall, accuracy, and precision, our proposed hybrid model outperforms several of the top standalone architectures, including ResNet50, VGG16, and a traditional ViT. The models in this framework are simple to understand, and the diagnostic accuracy is really good. This makes it an effective tool for supporting radiologists in making clinical decisions.

Authors

B. Sophia, B. Sanjai, S. Abinav, S. Jamaal Mohammed Aqmal
Kumaraguru College of Technology, India

Keywords

Hydrocephalus Classification, Hybrid Deep Learning, EfficientNet, Vision Transformer (ViT), Model Interpretability, LIME