Machine Learning Approaches to Improve Diagnosis and Management of Mammographic Calcifications

Abstract

Currently the most common and effective procedure of early detection of breast cancer is through screening mammography. Mammography detects not only invasive cancers, but also in situ lesions including ductal carcinoma in situ (DCIS), which is the most common non-invasive breast cancer. The focus of this dissertation is to investigate machine learning and deep learning technologies to improve detection and diagnosis of breast cancer from mammography, focusing on calcifications that represent a common type of suspicious lesions. This dissertation is composed of four main projects.The first project developed a transfer learning framework to predict occult invasive disease in DCIS by a machine learning approach. Performance was improved by using forced labeling and domain adaptation based upon data from two related diseases, atypical ductal hyperplasia (ADH) and invasive ductal carcinoma (IDC).The second project expanded the dataset fivefold, producing the largest cohort of DCIS cases to date. This rich dataset enabled rigorous studies to assess the performance and clinical utility for guiding patients with DCIS who are considering active surveillance or surgical planning.The third project developed a one-class anomaly detection system to detect calcifications with deep autoencoder network and dissimilarity index. Calcifications were detected by learning negative-only images. Using a very large mammography dataset, this study was one of the first in medical imaging to show model capacity changes based on different architectures as well as different number or type of training images.Finally, the fourth project broadened the diagnostic task to classify breast calcifications as benign or malignant using a deep multitask network. Unlike previous projects, algorithms here were designed to detect calcifications and classify cancers simultaneously. The multitask network model’s results showed improved performance compared to single classification networks. A correlation between two output losses’ weighting factors was also achieved by analyzing performance impacts from both weighting factors.In conclusion, this dissertation provides three main contributions. First, it provides a complete classification framework to distinguish invasive disease from DCIS based on the largest single institution dataset. Second, it demonstrates the potential of an anomaly detection approach to detect mammographic calcifications that may facilitate new clinical applications such as computer-aided triage and quality improvement. Third, a unified approach for multitask learning was used to perform detection and classification simultaneously, which provided the highest performance for classification from our group to date.