Author

Pradnya Patel

Date of Award

6-2021

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2021.

Degree Name

Master of Science (MS)

Department

Bioengineering

First Advisor

Yuling Yan

Abstract

Three deep learning models using convolutional neural network (CNN) were developed for the early detection of breast cancer and brain aneurysm. Model 1 was built for the detection of breast mass; it consists of 20 total layers including 5 convolutional layers, 5 maxpool layers with Rectifier Linear Unit as the activation function for feature extraction, one flatten layer, 4 batch normalization with fully connected layers, and one output layer. This CNN model was trained and validated on an open-source breast ultrasound dataset that contains a total of 830 images categorized into three classes: normal (133 images), benign (487 images), and malignant (210 images). The model was tested on our dataset collected at Santa Clara Valley Medical Center (SCVMC) consisting of 300 cases with breast masses of which 194 are benign (64.7%) and 106 were malignant (35.3%). The final accuracy of the model on this test set achieved is 94.89%.

Models 2 & 3 are built for the detection of brain aneurysms as shown in magnetic resonance angiography (MRA) images. The MRA dataset containing contiguous MRA images representing normal and aneurysms were provided by collaborator radiologists at SCVMC. Model 2 comprises of 11 layers (3 convolutional layers, 3 max pool layers, 3 fully connected layers, a softmax layer, and an output layer) and was trained and validated on a larger open-source medical dataset (CBIS-DDSM mammogram) with 90% of which used for the training and 10% for the validation of the model. A transfer learning approach was then used to retrain the model for the detection of aneurysm using the MRA dataset, which contains 29024 images for normal and 25245 images for aneurysm cases collected from 100 healthy subjects and 100 patients with aneurysm. Model 3 represents the retrained model, which achieved a test accuracy of 76.04%, this is a preliminary result for this study.

Download

Share

COinS