Abstract
The number of medical imaging studies is increasing at a disproportionate rate to the number of professional radiologist required to perform interpretation and diagnosis. As a result, radiologists become a source of a bottleneck in healthcare systems. The combination of the increasing pressure placed on radiologists, and the error and subjectivity inherent when interpreting medical images, result in a high number of misdiagnoses. To address the aforementioned issues, the use of computer aided detection and diagnosis systems have been proposed as a solution to aid clinicians. However, the early iteration of such systems which used manual and task-specific feature extraction techniques have yet to consistently match the sensitivity of professional radiologists, and tend to generate high false-positive rates. The use of deep learning methodologies has resulted in state-of-the-art performance on common computer vision tasks and demonstrated efficacy for performing radiological imaging analysis. In our dissertation, we aim to investigate the use of deep learning methodologies for performing the tasks of abnormality detection, classification, and segmentation in mammographic imaging. We also propose the use of meta-heuristic algorithms to fine tune our solutions without the need for external context, such as the neural network’s gradient information or knowledge of internal neuron connectivity. We compared the use of Genetic and Particle Swarm Optimization algorithms that used a shared neural network representation to fine tune the model trained using gradient descent and backpropagation. We used these meta-heuristic algorithms to search for the set of neural network weights that represented the global minimum solution. The abnormality detection and classification tasks were performed independently using ResNet50 and Xception architectures, while semantic segmentation was performed using a U-Net model. The Xception architecture out performed the ResNet baseline for both tasks, with the Xception baseline model achieving an Area Under the Curve (AUC) score of 0.72 and 0.77 on the classification and detection tasks respectively. The U-Net segmentation model struggled to achieve good metric performance with a sensitivity score of 27.42% and specificity score of 99.23%. Regarding the meta-heuristic algorithms, the PSO was predisposed to biasing the model to predict only negative cases, and for all tasks degraded the model’s performance...
M.Sc. (Computer Science)