Transfer Learning Enhanced CNN With GRAD-CAM for Early Alzheimer’s Detection on OASIS MRI

Abstract

Early and proper diagnosis of the Alzheimer disease (AD) has been a serious issue in the clinical practice, mainly because of the reticence nature of pre-symptomatic atrophy patterns and the burden of undiagnosed preliminary period dementia. This is a diagnostic gap that can have a strong influence on timely intervention and patient care outcomes. The objective of the study was to design and test an automated deep learning model to classify brain MRI images into one of the following categories; namely, an initial stage of dementia progression or cognitive normality, and to focus specifically on the issue of early detection and model interpretability. We used a transfer-learning method based on the use of state-of-the-art convolutional neural network (CNN) backbones, namely ResNet50 and EfficientNet-B3, which were pretrained with ImageNet weights. A special classification head was developed and connected to the backbone and extensive fine-tuning plans were undertaken. In order to offer interpretable information about the model decision-making, we also integrated explainability frameworks such as Grad-CAM++ and Integrated Gradients. On a test set that was held out and based on the OASIS dataset, the overall accuracy of the proposed framework was found to be 98.07% ±0.45%. Class-specific performance was very sensitive and specific with Cognitively Normal (CN) having a sensitivity of 99 and specificity of 97, and Demented (DEM) having a sensitivity of 96 and specificity of 98. The average F1-score was 97.8% standard deviation of 0.6, and the AUC-ROC of 0.995, 0.988 and 0.970 with CN, DEM and Very Demented (VD) respectively. The expected calibration error (ECE) was 0.03 which revealed good calibration probability estimations. The findings prove the feasibility of transfer-learning-based-enhanced CNNs to provide automated classification of dementia stages using structural MRI with high accuracy indicating that there is a high likelihood of its implementation in clinical screening pipelines.