Multi-Model Deep Learning for Dementia Detection: Addressing Data and Model Limitations

Areej Bayahya^1,2*fares Jammal¹Haneen Banjar¹Fathy Elbouraey Eassa¹Omar Talabay³Sultan Alamri⁴

• ¹Department of Computer Science, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah, Saudi Arabia, jeddah, Saudi Arabia
• ²Software Engineering Department, College of Engineering, University of Business and Technology (UBT), Jeddah 21448, Saudi Arabia, Jeddah, Saudi Arabia
• ³Center of Artificial Intelligence in Precision Medicines, King Abdulaziz University, Jeddah, Saudi Arabia, Jeddah, Saudi Arabia
• ⁴Department of Family Medicine, King Abdulaziz University, Jeddah 21589, Saudi Arabia, jeddah, Saudi Arabia

Multiple deep neural network architectures have significantly transformed the field of medical imaging, offering unprecedented capabilities for classification in sMRI. However, this study addressed the limitations of state-of-the-art deep learning models and difficulties encountered by preprocessing data to classify dementia diseases using sMRI images. This study examined the efficacy of eight pretrained CNNs models, as well as ViT, Multimodal attention and CapsNet to evaluate accuracy, specificity and sensitivity in the classification of dementia disease. This study was trained on 10000 images for each class, validated on 3000 images for each class and tested on 850 images in each class which are dementia, MCI, healthy control. The sMRI dataset balanced between all type of classification gained from ADNI 6 . It evaluated the performance of image classification in 2D grid as slices structural sMRI. Among the evaluated models, the 3D-CNN and Multimodal attention exhibited the highest metrics, achieving an accuracy of 84% and 86%respectively, specificity of 83% and 86% respectively, and sensitivity of 84% and 86% respectively. The outcomes showed diminished and biased to some classes in all models. The sensitivity of AD in ViT as well as CapsNet is measured at 100%. In contrast, the precision in ViT as well as CapsNet is achieved in AD 43% while the precision for other classifications is determined to be 0%.This study suggests that the models require improvements in data representation using computer vision methods, or model architecture to enhance overall performance. This study conducts an in-depth analysis of various algorithms, including transformer and caps network when used sMRI to classify dementia disease. It aimed to compare their performance and uncover underlying factors contributing to the suboptimal results in image classification and dementia diagnosis. These models, while effective in certain aspects, often face challenges in feature extraction, accuracy, and computational efficiency.