Automatic Classification, Detection and Segmentation of Breast Arterial Calcification on Digital Mammography Images Using Deep Learning

Maguire, Dominic

Abstract

Cardiovascular disease (CVD) is the leading cause of premature death in the United Kingdom with one type, coronary artery disease, killing more than twice as many women as breast cancer. Conventional CVD risk factors have been shown to have less accuracy for females who are considered low-risk. Recently, researchers have noted that breast arterial calcification (BAC), which is regularly observed as an incidental finding on mammograms, could be used to risk-stratify women for CVD.

In 2023, almost 2 million women attended breast screening clinics in England. Automatic BAC detection on mammograms could provide vital additional cardiovascular information, without the need for further invasive tests or radiation exposure, and could direct patients to relevant clinical pathways or therapies.

As a first step in automating the BAC grading process, I developed deep learning models for BAC classification, object detection and segmentation using an anonymised dataset which was annotated for the presence and location of BAC under the guidance of two consultant radiologists. Data augmentation was used in both the classification and object detection networks, increasing the training data size.

My modified ResNet22 network showed promise in classifying the presence or absence of BAC at image level, attaining a test accuracy of 80%, indicating that this method could be used as a simple flag for this purpose. I also used this network for feature extraction in Faster R-CNN and YOLO BAC object detection models. Despite improving on a recent similar study, these latter networks performed poorly with very low average precision scores at several thresholds. As an improvement, this study developed a DeepLabv3+-based BAC segmentation network which doubled the IoU obtained by another study using a similar model and achieved a BFScore of over 70% specifically for BAC.

Based on the findings of this research, a two-step pipeline is recommended with our classifier triaging mammographic images for BAC and our segmentation model providing an indication of the extent of its presence. This could provide the basis for further research in order to realise the potential of concurrent, automatic BAC grading for women undergoing mammographic imaging.

• Thesis Type: Thesis
• Online Publication Date: Jan 23, 2025
• Deposit Date: Nov 20, 2024
• Publicly Available Date: Feb 24, 2025
• Award Date: Jan 23, 2025