An efficient deep learning approach to detect multiple neurodegenerative diseases using image data

Abstract

In today’s time early and accurate diagnosis of the neurodegenerative diseases (such as Alzheimer Disease (AD), Parkinson Disease (PD), Frontotemporal Dementia (FTD)) is essential to timely intervention and is quite laborious since the Magnetic Resonance Imaging (MRI) scans of all of them carry same set of characteristics. The deep learning pipeline proposed in this study will utilize 2D sagittal sectioning of 3D MRI columns and will be focusing on such crucial regions as the hippocampus, substantia nigra, and forehead lobe. These pictures were downsized to the identical standard of 224x224x3 and heavily augmented to promote generalization. We benchmarked the transfer learning classification performance of a number of pre-learning CNN models. Among the standard models, MobileNetV2 exhibited superior performance on the test set (94.36%), as opposed to EfficientNetB0 (85.47%), ResNet50 (68.29%), DenseNet121 (62.65%), and VGG19 (33.33%). However, MobileNetV2 exhibited a little overfitting. To address the challenges of accuracy problem of multiclass disease detection and overfitting issue we proposed SadNetV1 that was constructed by integrating MobileNetV2, Squeeze-and-Excitation (SE) blocks and Spatial Attention mechanism to facilitate effective dimensionality reduction in capturing channel and spatial dependence. The proposed SadNetV1 model showed improved performance of 96.15% test, 96.84% train, and 97.11% validation accuracy and can offer generalization in complex MRI images and could differentiate between AD and PD, and FTD. These results give credence to the potential of attention-augmented lightweight networks in effective categorization of neurodegenerative diseases in clinics.