The impact of the orientation of MRI slices on the accuracy of Alzheimer’s disease classification using convolutional neural networks (CNNs)

Background: Alzheimer’s disease (AD) is the leading cause of major neurocognitive disorders, affecting approximately 50 million people worldwide. Due to its high prevalence, AD significantly impacts patients’ quality of life and poses a substantial challenge to healthcare systems. Diagnosis is intricate, with specificity and sensitivity rates falling below the ideal. Early identification of AD is essential to increase the effectiveness of pharmacotherapeutic treatment and improve quality of life. Consequently, there is a quest for innovative methods, such as machine learning and deep learning, to automate the diagnosis of AD in its early stages. Methods: We developed and validated a convolutional neural network (CNN) algorithm using the Keras Sequential API in Python to investigate the impact of slicing T1-weighted magnetic resonance images on the classification of patients with mild cognitive impairment (MCI) and healthy patients (NC), grouped based on scores on the Mini-Mental State Examination (MMSE). We selected 318 patients (250 healthy and 68 MCI) with a minimum of 16 years of education (equivalent to a completed undergraduate degree). The training, testing, and validation datasets were split in a 70/15/15 ratio for each slice. Results: The CNN achieved high accuracy values in classifying healthy and MCI groups, ranging between 97% and 99% depending on the slice, the number of training epochs, and batch size. In addition to precision, the F1-score, recall, and precision parameters were also evaluated, with values above 91%. Generally, the coronal slice produced the best results, followed by the axial and the sagittal slices, which nevertheless showed high performance, standing out individually in different evaluation parameters. Notably, the choice of batch size and the number of epochs also influenced the network’s classification. Conclusions: Our study findings indicate that utilizing CNN in conjunction with selecting a coronal slice proves to be a promising tool for facilitating the early-stage diagnosis of neurodegenerative diseases, such as AD, through magnetic resonance imaging analysis, enabling more effective treatments and appropriate future planning. Moving forward, we aim to investigate whether these results replicate across other imaging modalities, such as positron emission tomography, and explore additional datasets.