Identification of Malignant Attributes in Breast Ultrasound using a Fully Convolutional Deep Learning Network and Semantic Segmentation

Breast ultrasonography is a non-radiation imaging technology that is used to detect and categorize breast malignancies. Individuals have a minimal prevalence of adverse symptoms to it, and it is simple to incorporate therapeutic procedures. The proposed study has used a fully convolutional deep learning network to extract features from breast ultrasound images in order to identify different characteristics of breast imaging and data system terminology. This in turn simplify the procedure of categorizing tumors as benign or malignant. Using the BI-RADS vocabulary, 378 breast ultrasound images are analyzed and from that seven potentially cancerous characteristics are discovered. The mean accuracy and mean IU for the feature extraction are 32.82% and 28.88% respectively. Both the area under the curve and the normalized interconnection over union were observed to be higher than the results obtained by the equivalent feature extraction connections such as SegNet and U-Net by using the same range of data. The area under ROC curve has been 89.47% as well as the graded interconnection placed above a white confederation was 85.35%. By utilizing the ultrasonography to screen for breast cancer, the proposed research study suggests that it would be beneficial to make use of a deep learning network in conjunction with the BI-RADS terminology as a substantial supplementary diagnostic tool.