A Novel Deep Learning Algorithm for Classification of Dementia Subtypes

Health centers capture enormous amounts of image diagnostic data at a pace far surpassing what traditional methods of analysis can process. Deep learning is one of the best ways to make predictions based on large and heterogeneous datasets. This is particularly pertinent for images obtained with different magnetic resonance imaging scanning parameters across health centers, given the underlying flexibility of these models. Deep learning avoids biases in data collection and increases our capacity for multi-site image harmonization. Current models may provide appealing, complementary, and feasible solutions as decision support tools for diagnostic purposes, particularly when relying solely on brain imaging data, as is common in clinics within Latin American Countries (LAC). In doing so, they provide an alternative solution to costly biomarkers that are only available in high-income countries.

In this pilot project, I propose to develop a Deep Learning algorithm to extract the most important brain image features that are invariant across health-centers for classifying between patients with Alzheimer’s Disease and patients with behavioral variant Frontotemporal Dementia. Then I will test the algorithm in research and clinical datasets obtained in LAC to assess the generalizability of its predictions. Unlike other traditional machine learning approaches that are out of reach of LAC because of the cost and unavailability of specialized personnel, the specific goal of this pilot project is to create and optimize a more advanced deep learning algorithm based on magnetic resonance imaging automatic feature selection, while validating its diagnostic accuracy on heterogeneous data obtained in different health centers from LAC. As the next step after this pilot project, the long-term goal is to produce an app for clinical use based on the developed algorithm.