Breast lesion detection and visualization utilizing artificial intelligence and the H-scan

W e incorporated raw ultrasound parameters into artificial intelligence-based breast cancer diagnosis to achieve improved accuracy compared to radiologists and deep learning (DL). 76 patients with suspicious breast lesions were ultrasound-imaged using a Samsung RS85 system equipped with a 9.4 MHz center frequency transducer. The patients underwent biopsy, and the biopsy results were used as a reference gold standard: n=53 for benign and n=23 for malignant. Ten radiologists reviewed the ultrasound images and provided BI-RADS (Breast Imaging Reporting and Data System) scores. A previously trained DL product with a modified fully convolutional network and GoogLeNet contoured the breast lesion boundaries. Within the contoured lesions, ultrasound parameters were extracted from the radiofrequency, envelope, and log-compressed data: (1) H-scan color level, (2) lesion boundary shape using convex hull, (3) B-mode boundary standard deviation (STD), (4) B-mode STD, and (5) Burr distribution b. To quantify breast condition, multiparametric analysis combining the 5 features was performed using principal component analysis, resulting in the first principal component (PC1). The PC1 outputs within a lesion were overlaid on B-mode images. We calculated the area under the curve (AUC) to evaluate performance. We compared AUC results from radiologists, the DL product, and our PC1 quantification. The PC1 showed the highest AUC. Further, utilizing this PC1 quantification, we visualized the localized probability of malignancy, illustrating BI-RADS score differences using a color display. Overall, we demonstrated the potential of utilizing raw ultrasound parameters to improve DL performance and to achieve higher diagnostic accuracy than radiologists for breast cancer.