Available at: https://digitalcommons.calpoly.edu/theses/3182
Date of Award
11-2025
Degree Name
MS in Engineering - Bioengineering
Department/Program
Biomedical Engineering
College
College of Engineering
Advisor
Robert Szlavik
Advisor Department
Biomedical Engineering
Advisor College
College of Engineering
Abstract
Breast cancer is the second most common form of cancer and often goes undetected in its initial stages due to its subtle symptoms. As the tumors grow, they become more difficult to surgically remove with clean borders. To minimize the chance of recurrence, a 2D convolutional neural network is developed in this work for delineating tumor boundaries. Specifically, the U-shaped network model is trained, validated, and tested with longitudinal MRI scans of patients diagnosed with breast cancer and undergoing neoadjuvant therapy. For training, image masks were generated as ground truths using signal enhancement ratio segmentation, thresholding, and contour detection. The effectiveness of two lightweight models, with and without batch normalization and dropout layers, were compared for accuracy, loss, and processing times. The model including batch normalization and dropout layers outperformed the network without these functions; the enhanced model was additionally compared to several other more intricate machine learning algorithms for tumor-feature detection. The convolutional neural network designed for this work performed with higher accuracy and lower loss rates than the majority of the complex models, demonstrating that algorithm simplicity does not mean reduced functionality of the machine learning system.