Date of Award
Summer 2023
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Povinelli, Richard J.
Second Advisor
Ye, Dong Hye
Third Advisor
Yu, Bing
Abstract
Breast-conserving surgery (BCS) is a widely used treatment for breast cancer, but ensuring the complete removal of cancer cells from the surgical margins remains a challenge. Deep ultraviolet (DUV) fluorescence scanning microscopy offers a potential solution by providing real-time whole-surface imaging of resected tissues during BCS. However, interpreting DUV images for margin assessment requires an automated classification method. This dissertation addresses this need by proposing a deep learning-based classification approach for DUV fluorescence images in intra-operative margin assessment of breast cancer.To overcome the limited availability of DUV image datasets and potential over- fitting, the study combines patch-level classification using transfer learning with regional importance maps generated through the Grad-CAM++ algorithm. The proposed method- ology involves dividing DUV whole-slide images into smaller patches, converting them to grayscale, and analyzing pixel values to identify valid patches. A pre-trained ResNet50 network and an XGBoost classifier are utilized for patch-level classification, while Grad- CAM++ generates regional importance maps for the entire DUV image. The decision fusion method combines patch-level classification labels and regional importance maps to determine the whole-slide image (WSI)-level classification label by calculating the total number of malignant patches and comparing it to a threshold percent- age of total foreground patches. A binary classification is obtained for the entire WSI. The proposed methodology is implemented using PyTorch and a dataset consisting of 60 DUV images of breast tissue samples. The DUV images were obtained using a custom DUV-Fluorescence Scanning Microscopy system, which provided high-resolution images with fluorescence staining for accurate tissue classification. The results of this study contribute to the field of intra-operative margin assessment in breast cancer by demonstrating the effectiveness of deep learning-based classification of DUV images. The combination of transfer learning, regional importance maps, and deci- sion fusion provides a robust approach for accurately classifying breast tissue as malignant or normal/benign. This research opens new avenues for utilizing deep learning techniques in DUV fluorescence imaging and has the potential to improve surgical outcomes in breast- conserving surgery.