A Deep-Learning System for Detecting the Brain MRI Anatomical Plane to be Examined with Priority in Alzheimer's Disease

Brain Magnetic Resonance Imaging (MRI) is an effective technology to catch and analyse alterations of the brain morphology. An overall MRI view in all anatomical planes provides the most comprehensive information, however it requires attention and time. In addition, the choice of the best diagnostic anatomical plane depends on the location and shape of the specific brain structures. Moreover, each radiologist selects the best strategy to obtain diagnostic information based on representativeness, anchoring, and availability, resulting highly subjective. In this context, a system designed to automatically detect the most informative anatomical plane based on the brain structure alterations can be useful in guiding, thus accelerating the interpretation of brain MRI information, especially in those cases where MRI is among the primary means of producing clinical diagnoses, such as in neurodegenerative diseases. Currently, there are not solutions in the field of artificial intelligence that are specifically designed to fulfill this task. Therefore, the aim of this research is to investigate the atypicality of the brain structures involved the most in Alzheimer's Disease (AD) to prioritize the reading of the anatomical plane with the highest ability to detect morphological changes related to AD. To this aim, we implemented a deep-learning system in which three lightweight and self-excluding classification models were utilized on MRI volumes of AD, Mild Cognitive Impairment (MCI), and cognitively normal individuals taken from the ADNI database. The first model is a Convolutional Long Short-Term Memory (ConvLSTM)-based neural network, whereas the others are two time-distributed convolutional neural networks combined with a ConvLSTM-based module each. According to our results, the best performing model classifies AD subjects with AUROC mean value of 99% from the axial plane. As for MCI subjects, the same ConvLSTM-based neural network classifies them with AUROC mean value of 96% from the sagittal plane. All models agree in indicating the axial and sagittal planes as the ones to examine first in the majority of AD and MCI subjects, respectively. The developed system has practical utility for subsequent segmentation of the brain structures of interest. Moreover, we firmly believe on its greatest potential if integrated in the MRI machine with the scope of prioritizing the display of the most informative anatomical plane on a subject-wise perspective.