Image Modality Classification of Stroke MRIs Using Deep Learning

Michael Salamone

Mentor: Dr. Peter Turkeltaub, Professor of Neurology and Rehabilitation Medicine; Director, Cognitive Recovery Laboratory; Director, Neuroscience of Language Training Program, Georgetown University Medical Center; Director, Aphasia Clinic, MedStar National Rehabilitation Network; Dr. Kyle Shattuck, SOM Research Track, Instructor, Rehabilitation Medicine, Georgetown University.

Date/Time: August 22nd, 2025 at 11:30 AM.

Abstract: Magnetic resonance imaging (MRI) modalities are widely used to support research in stroke diagnosis. Different modalities provide complementary information essential for accurate assessment. However, the amount of magnetic resonance data generated from different modalities can be difficult to manage manually, slowing analysis. This project addresses that challenge by developing an automated deep learning pipeline to classify brain MRI modalities in the Acute Stroke Imaging Database (AStrID), a clinical database dedicated to stroke imaging. AStrID curates brain MRIs from acute ischemic stroke patients to improve diagnostic accuracy, which enables faster treatment. This pipeline processes 75,000 midslices extracted from 3D NIfTI volumes into an HDF5 format, applying a preprocessing workflow. A 2D convolutional neural network (CNN) is implemented in PyTorch, iteratively refining the dataset. Misclassified images are manually reviewed for ground truth and precision, then relabeled or excluded before retraining. Using 10-fold cross-validation, the final model achieves a high classification accuracy across all modalities. This work provides a reproducible approach to modality classification in real-world clinical scenarios and represents a technical advancement in the AStrID pipeline, supporting broader efforts in automated stroke imaging detection and analysis.