Available at: https://digitalcommons.calpoly.edu/theses/2006
Date of Award
MS in Computer Science
The job of classifying or annotating brain tumors from MRI images can be time-consuming and difficult, even for radiologists. To increase the survival chances of a patient, medical practitioners desire a means for quick and accurate diagnosis. While datasets like CIFAR, ImageNet, and SVHN have tens of thousands, hundreds of thousands, or millions of samples, an MRI dataset may not have the same luxury of receiving accurate labels for each image containing a tumor. This work covers three models that classify brain tumors using a combination of convolutional neural networks and of the concept of capsule layers. Each network utilizes a hit-or-miss capsule layer to relate classes to capsule vectors in a one-to-one relationship. Additionally, this work proposes the use of deep active learning for picking the samples that can give the best model, PSP-HitNet, the most information when adding mini-batches of unlabeled data into the master, labeled training dataset. By using an uncertainty estimated querying strategy, PSP-HitNet approaches the best validation accuracy possible within the first 12-24% of added data from the unlabeled dataset, whereas random choosing takes until 30-50% of the unlabeled to reach the same performance.