Automatic deep-learning classification models for breast lesions

S. Hasan, A. Hasan
Princeton Day School,
United States

Keywords: machine learning, AI, classification, breast cancer, support vector machine, segmentation, characterization, classifier


Breast cancer vulnerability in women typically increases until they are about 74 years old, and then begins to steadily decline. As of 2019, an estimated 268,600 new cases of invasive breast cancer will be diagnosed among women. While 128.5 out of 100,000 new women are diagnosed with this cancer each year, 20.3 of these women eventually die of the disease. In addition, approximately 41,760 women and 500 men are expected to die from breast cancer this year. Stage 1 breast cancer means that the cancer cells have not spread to other areas from the area. Garra et al (1997) discovered that cancerous lesions can be distinguished from benign lesions using a size difference. The results in their paper demonstrated that invasive ductal carcinoma lesions appear larger in strain images compared to conventional B-mode ultrasound images whereas the sizes were approximately the same for benign lesions. Furthermore, malignant lesions exhibited higher contrast than benign lesions. However, size difference was a stronger classification feature. Barr et al (2012) performed a larger study (528 female patients with 635 lesions). In his study, if the ratio was higher than 1.0, the lesion was considered malignant. If the ratio was lower than 1.0, the lesion was considered to be benign. This project is based on the work of Garra et al and Barr et al to automate the classification using these criteria. Diagnosis by radiologists is highly dependent on elastography expertise whereas our algorithm is objective and can help guide a clinician to make an improved diagnosis. In our study, we determined the effectiveness of the apparent lesion-size discrepancy determined by automatic segmentation algorithms between the processed ultrasound B-Mode and strain image of in vivo patient data as a classification criterion between benign fibroadenoma and adenocarcinoma. Furthermore, we designed support-vector machine classifiers based on lesion-size discrepancy, intensity contrast in the elasticity image, and intensity contrast in the B-Mode image, intensity contrast discrepancy to be effective classification models by which to differentiate between fibroadenoma and adenocarcinoma.