In recent years, Deep Learning (DL) has shown promising results with regard to conducting AI tasks such as computer vision and speech recognition. Specifically, DL demonstrated the state-of-the-art in computer vision tasks including image classification, segmentation, localization, and annotation. Convolutional Neural Network (CNN) models in DL have been applied to prevention, detection, and diagnosis in predictive medicine. Image segmentation plays a significant role in predictive medicine. However, there are huge challenges when performing DL-based automatic segmentation due to the nature of medical images such as heterogeneous modalities and formats, the very limited labeled training data, and the high-class imbalance in the labeled data. Furthermore, automatic segmentation becomes a challenging task, especially for Magnetic Resonance Images (MRI). In reality, it is a time- consuming procedure that requires trained biomedical experts to manually segment or annotate such MRI datasets. The need for automated segmentation or annotation is what motivates our work.

In this thesis, we propose a semi-automated approach that aims to segment the claustrum in brain MRI images. We recognize that the claustrum is an information hub of human brains and can be used to find significant patterns from the segmentations. We applied a 2-Dimensional CNN model called U-net to segment the human brain dataset comprising 30 manually annotated subjects provided to us by the Department of Psychiatry at the University of Missouri-Kansas City. Our approach consisted of the following steps: (1) preprocessing, including converting, the data into Digital Imaging and Communications in Medicine (DICOM), re-sampling and selecting the claustrum slices, and applying an ROI selection; (2) building the claustrum model; (3) automatic segmentation; and (4) evaluation and validation. For the model validation, we used the cross-validation technique with n = 5. We administered the Dice coefficient index to evaluate the results and we achieved a Dice score of approximately 70%. A domain expert also evaluated the results.