A deep learning system for classifying structural brain MRIs in elderly patients in a low-resource clinical setting

Purpose: With the expected doubling of people aged 60 and older by 2050, the understanding of age-related structural brain changes (SBC) remains minimal, particularly in low- and middle-income countries (LMIC). Furthermore, radiologists and radiology resources in LMIC are overloaded due to the scarcity of trained experts. Several studies demonstrated the potential of using artificial intelligence to support radiologists in identifying SBC on Magnetic Resonance Imaging (MRI). Nevertheless, inaccurately reporting clinically significant SBC in the elderly as “normal for age” remains a challenge. Furthermore, most of the elderly neuroimaging datasets come from western populations, leaving other populations with an unmet need for data-driven solutions that work for their patients and affordable machines. To this end, we develop and validate a deep learning decision support system to classify brain aging using MRIs from elderly patients that might require clinical interventions in a patient population from Palestine.

Methods/Materials: The dataset is composed of retrospective brain MRIs for 70 patients above the age of 50, collected from a multi-specialty outpatient medical center using an open-MRI 0.4T machine. Patients were labeled based on the radiology findings and clinical presentation recorded in the PACs system. The dataset was split into train, validation, and testing, where data preprocessing and augmentation were applied before training the Convolutional Neural Network-based system.

Results: The system achieved an Area Under Receiving Operating Curve (AUROC) of 0.738 and an Area Under the Precision-Recall Curve (AUPRC) of 0.772. To the best of our knowledge, this work is the first to demonstrate the feasibility and potential of using data collected from open-MRI, for understudied populations from the Middle East.

Conclusions: Our system aims to contribute to the understanding of aging from brain MRIs, optimizing the radiologist's time, and ultimately help initiate targeted interventions to reduce risks and improve the patient’s quality of life. In future work, we aim to investigate the system's ability to predict specific clinical findings and perform tasks that support radiologists in their practice.