Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Interdisciplinary and Graduate Studies

Degree Program

Interdisciplinary Studies (Individualized Degree), PhD

Committee Chair

El-Baz, Ayman

Committee Co-Chair (if applicable)

Giridharan, Guruprasad

Committee Member

Giridharan, Guruprasad

Committee Member

Frieboes, Hermann

Committee Member

Mohamed, Tamer

Committee Member

Gondim, Dibson

Author's Keywords

Deep learning; medical imaging


Machine learning, a sub-discipline in the domain of artificial intelligence, concentrates on algorithms able to learn and/or adapt their structure (e.g., parameters) based on a set of observed data. The adaptation is performed by optimizing over a cost function. Machine learning obtained a great attention in the biomedical community because it offers a promise for improving sensitivity and/or specificity of detection and diagnosis of diseases. It also can increase objectivity of the decision making, decrease the time and effort on health care professionals during the process of disease detection and diagnosis. The potential impact of machine learning is greater than ever due to the increase in medical data being acquired, the presence of novel modalities being developed and the complexity of medical data. In all of these scenarios, machine learning can come up with new tools for interpreting the complex datasets that confront clinicians. Much of the excitement for the application of machine learning to biomedical research comes from the development of deep learning which is modeled after computation in the brain. Deep learning can help in attaining insights that would be impossible to obtain through manual analysis. Deep learning algorithms and in particular convolutional neural networks are different from traditional machine learning approaches. Deep learning algorithms are known by their ability to learn complex representations to enhance pattern recognition from raw data. On the other hand, traditional machine learning requires human engineering and domain expertise to design feature extractors and structure data. With increasing demands upon current radiologists, there are growing needs for automating the diagnosis. This is a concern that deep learning is able to address. In this dissertation, we present four different successful applications of deep learning for diseases diagnosis. All the work presented in the dissertation utilizes medical images. In the first application, we introduce a deep-learning based computer-aided diagnostic system for the early detection of acute renal transplant rejection. The system is based on the fusion of both imaging markers (apparent diffusion coefficients derived from diffusion-weighted magnetic resonance imaging) and clinical biomarkers (creatinine clearance and serum plasma creatinine). The fused data is then used as an input to train and test a convolutional neural network based classifier. The proposed system is tested on scans collected from 56 subjects from geographically diverse populations and different scanner types/image collection protocols. The overall accuracy of the proposed system is 92.9% with 93.3% sensitivity and 92.3% specificity in distinguishing non-rejected kidney transplants from rejected ones. In the second application, we propose a novel deep learning approach for the automated segmentation and quantification of the LV from cardiac cine MR images. We aimed at achieving lower errors for the estimated heart parameters compared to the previous studies by proposing a novel deep learning segmentation method. Using fully convolutional neural networks, we proposed novel methods for the extraction of a region of interest that contains the left ventricle, and the segmentation of the left ventricle. Following myocardial segmentation, functional and mass parameters of the left ventricle are estimated. Automated Cardiac Diagnosis Challenge dataset was used to validate our framework, which gave better segmentation, accurate estimation of cardiac parameters, and produced less error compared to other methods applied on the same dataset. Furthermore, we showed that our segmentation approach generalizes well across different datasets by testing its performance on a locally acquired dataset. In the third application, we propose a novel deep learning approach for automated quantification of strain from cardiac cine MR images of mice. For strain analysis, we developed a Laplace-based approach to track the LV wall points by solving the Laplace equation between the LV contours of each two successive image frames over the cardiac cycle. Following tracking, the strain estimation is performed using the Lagrangian-based approach. This new automated system for strain analysis was validated by comparing the outcome of these analysis with the tagged MR images from the same mice. There were no significant differences between the strain data obtained from our algorithm using cine compared to tagged MR imaging. In the fourth application, we demonstrate how a deep learning approach can be utilized for the automated classification of kidney histopathological images. Our approach can classify four classes: the fat, the parenchyma, the clear cell renal cell carcinoma, and the unusual cancer which has been discovered recently, called clear cell papillary renal cell carcinoma. Our framework consists of three convolutional neural networks and the whole-slide kidney images were divided into patches with three different sizes to be inputted to the networks. Our approach can provide patch-wise and pixel-wise classification. Our approach classified the four classes accurately and surpassed other state-of-the-art methods such as ResNet (pixel accuracy: 0.89 Resnet18, 0.93 proposed). In conclusion, the results of our proposed systems demonstrate the potential of deep learning for the efficient, reproducible, fast, and affordable disease diagnosis.