Breast Cancer Detection and Sub-typing Using Subtraction of Temporally Sequential Mammograms and Machine Learning: Assessment of Invasiveness and Tumor Grade.

Despite significant advancements, Breast Cancer (BC) remains a leading cause of morbidity and mortality among women worldwide. The management of BC, as well as the patient prognosis, depends on the tumor invasiveness and grade. Definitive classification of both is provided by biopsy and histopathology. However, this process is both invasive and increases delays in the diagnosis. In this study, an algorithm is developed for the automatic detection and sub-typing of BC type as in situ vs. invasive and tumor grading as grade 2 vs. grade 3 from mammographic images and other patient data. Subtraction of temporally sequential digital mammograms and feature-based Machine Learning (ML) were combined to achieve this goal. The methodology involves two main steps: (1) detection of Regions of Interest (ROIs) and (2) classification of the ROIs. The algorithm was developed using a new dataset of 164 images, with precise annotations. Ninety-six image features and 12 epidemiological and personal history features were collected. Eight feature selection algorithms and 10 classifiers were evaluated for identifying the best models. The algorithm achieved 91.2% accuracy and 0.91 AUC for in situ vs. invasive classification, and 92.8% accuracy and 0.92 AUC for tumor grading. These are the first such results reported in the literature. By reducing the reliance on biopsies, the algorithm provides a faster, non-invasive, and accurate tool for the subtyping of BC. When translated to clinical practice, it has the potential to enhance diagnostic efficiency, improve patient outcomes, and lower BC-related mortality rates.