In this Ph.D. thesis, we develop effective techniques for medical image classification, particularly, for histopathological and magnetic resonance images (MRI). Our techniques are capable of handling the high variability in the content of such images. Handcrafted techniques based on texture analysis are used for the classification task. We also use deep learning models but training such models from scratch can be a challenging process, instead, we employ deep features and transfer learning. First, we propose a combined texture-based feature representation that is computed in the complex shearlet domain for histopathological image classification. With complex coefficients, we examine both the magnitude and relative phase of shearlets to form the feature space. Our proposed techniques are successful for histopathological image classification. Furthermore, we investigate their ability to generalize to MRI datasets that present an additional challenge, namely high dimensionality. An MRI sample consists of a large number of slices. Our proposed shearlet-based feature representation for histopathological images cannot be used without adjustment. Therefore, we consider the 3D shearlet transform given the volumetric nature of MRI data. An advantage of the 3D shearlet transform is that it takes into consideration adjacent slices of MRI data. Secondly, we study the classification of histopathological images using pre-trained deep learning models. A pre-trained deep learning model can act as a starting point for datasets with a limited number of samples. Therefore, we used various models either as unsupervised feature extractors, or weight initializers to classify histopathological images. When it comes to MRI samples, fine-tuning a deep learning model is not straightforward. Pre-trained models are trained on RGB images which have a channel size of 3, but an MRI sample has a larger number of slices. Fine-tuning a convolutional neural network (CNN) requires adjusting a model to work with MRI data. We fine-tune pre-trained models and then use them as feature extractors. Thereafter, we demonstrate the effectiveness of fine-tuned deep features with classical machine learning (ML) classifiers, namely a support vector machine and a decision tree bagger. Furthermore, instead of using a classical ML classifier for the MRI sample, we built a custom CNN that takes both the 3D shearlet descriptors and deep features as an input. This custom network processes our feature representation end-to-end and then classifies an MRI sample. Our custom CNN is more effective in comparison to a classical ML on a hidden MRI dataset. It is an indication that our CNN model is less susceptible to over-fitting.

Polymicrogyria (PMG) is one brain disease that mainly occurs in the pediatric brain. Heavy PMG will cause seizures, delayed development, and a series of problems. For this reason, it is critical to effectively identify PMG and start early treatment. Radiologists typically identify PMG through magnetic resonance imaging scans. In this study, we create and open a pediatric MRI dataset (named PPMR dataset) including PMG and controls from the Children's Hospital of Eastern Ontario (CHEO), Ottawa, Canada. The difference between PMG MRIs and control MRIs is subtle and the true distribution of the features of the disease is unknown. Hence, we propose a novel center-based deep contrastive metric learning loss function (named cDCM Loss) to deal with this difficult problem. Cross-entropy-based loss functions do not lead to models with good generalization on small and imbalanced dataset with partially known distributions. We conduct exhaustive experiments on a modified CIFAR-10 dataset to demonstrate the efficacy of our proposed loss function compared to cross-entropy-based loss functions and the state-of-the-art Deep SAD loss function. Additionally, based on our proposed loss function, we customize a deep learning model structure that integrates dilated convolution, squeeze-and-excitation blocks and feature fusion for our PPMR dataset, to achieve 92.01% recall. Since our suggested method is a computer-aided tool to assist radiologists in selecting potential PMG MRIs, 55.04% precision is acceptable. To our best knowledge, this research is the first to apply machine learning techniques to identify PMG only from MRI and our innovative method achieves better results than baseline methods.

Computed Tomography (CT) is often used for building 3D biomechanical models of human anatomy. This method exposes the subject to a significant x-ray dose and provides limited soft-tissue contrast. Magnetic Resonance Imaging (MRI) is a potential alternative to CT for this application, as MRI offers significantly better soft-tissue contrast and does not expose the subject to ionizing radiation. However, MRI requires long scan times to achieve 3D images at sufficient resolution, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). These long scan times can make subject motion a problem. This thesis describes my work to reduce scan time while achieving sufficient resolution, SNR, and CNR for 3D biomechanical modeling of (1) the human larynx, and (2) the human hip. I focused on two important strategies for reducing scan time and improving SNR and CNR: the design of RF coils optimized to detect MRI signals from the anatomy of interest, and the determination of MRI relaxation properties of the tissues being imaged (allowing optimization of imaging parameters to improve CNR between tissues). Work on the larynx was done in collaboration with the Thomson group in Mechanical Engineering at BYU. To produce a high-resolution 3D image of the larynx, a 2-channel phased array was constructed. Eight different coil element designs were analyzed for use in the array, and one chosen that provided the highest Q-ratio while still meeting the mechanical constraints of the problem. The phased array was tested by imaging a pig larynx, a good substitute for the human larynx. Excellent image quality was achieved and MR relaxometry was then performed on tissues in the larynx. The work on the hip was done in collaboration with the Anderson group in orthopedics at the University of Utah, who are building models of femoral acetabular impingement (FAI). Accurate imaging of hip cartilage requires injection of fluid into the hip joint capsule while in traction. To optimize contrast, MR relaxometry measurements were performed on saline, isovue, and lidocaine solutions (all typically injected into the hip). Our analysis showed that these substances actually should not be used for MR imaging of the hip, and alternate strategies should be explored as a result.