Tohoku University reports that automated analysis of breast ultrasounds using SAS deep learning supports accurate and low-burden cancer diagnostics

Published article reveals findings on breast ultrasounds utilizing SAS Viya

Research from Tohoku University School of Medicine has applied SAS® deep learning software in breast cancer diagnostics to enable highly accurate classification of masses in breast ultrasound images. The research was published in Physics in Medicine & Biology1 and SAS Japan supported the study through software, equipment and technological assistance as a part of the company’s efforts to help academic researchers and foster analytics talent at colleges and universities. It also exemplified the company’s involvement in the Data for Good movement, which focuses on using data and analytics to help humanity and improve society.

A mammogram is commonly used for diagnosing breast cancer. However, it does not provide sufficient accuracy in women with dense breast tissue. In such a case, breast ultrasound (sonography) is used. Interpretations of ultrasound images are based on subjective opinions along with the experience of radiologists and physicians. In many cases, additional invasive biopsy is required because of high false-positive rates (rates of benign masses diagnosed as malignant), which results in higher physical and psychological burdens for patients. AI-based imaging helps guide diagnosis and enables more accurate breast mass classification.

Convolutional neural network (CNN), a deep learning method allowing highly accurate interpretation of diagnostic images, automatically learns a large volume of features required for effective identification; thus, helping find breast mass characteristics that humans cannot detect. In this study, researchers adopted ensemble learning, an approach that makes highly accurate identification possible by connecting the output of VGG19 and ResNet152, types of CNN models. SAS® Viya® has been designed for users to implement complicated models with simple codes using SAS, Python and R programming languages. This has enabled medical students to build various models and assess the accuracy.

As the data set used in this study contained multiple images of each patient, researchers adopted an approach to classify masses by patient, not by image. This resulted in a highly accurate classification model with a 90.9% sensitivity (rate of appropriately classified malignant masses), 87.0% specificity (rate of appropriately classified benign masses) and 0.951 AUC, a common criterion for evaluating the accuracy of AI (machine learning). Based on a heat map analysis for evaluating what areas in images the CNN model looks at in providing classification results, researchers found that there are potential areas to look at outside masses for classification. This suggests new viewpoints for diagnostic imaging. SAS Viya allows researchers to build models flexibly while incorporating AI technology and to collect the evaluation scores they need to include in scientific articles, where statistical accuracy is essential. Furthermore, AI, which tends to be a black box, explains how the result was obtained.

"With SAS' support, we applied deep learning to breast ultrasounds and published an article about a highly accurate computer-aided diagnostic system," said Professor Takuhiro Yamaguchi. "This system still has some challenges such as not being able to clearly show the process of classification. When this is clarified and the system is put into practical use in the clinical setting, it is expected to reduce the burdens of physicians and patients and also lower medical costs. To that end, we will further our research in cooperation with SAS."

AI is a promising technology that is expected to help solve challenges in diverse areas. Tohoku University School of Medicine adopts cutting-edge technologies in life sciences, shares research findings across academic fields and disciplines, and keeps producing talent. The SAS academic initiative team that took charge of this project continues to broaden support for research and education at educational institutions, aiming to generate professionals who use analytics for various research and business applications.

About Tohoku University School of Medicine

Founded by the feudal domain, Sendai-Han, Tohoku University School of Medicine began as a Sendai-Han Medical School in 1817. At the same time, a free dispensary (what we might call a postgraduate school) was built. After it was reorganized as Miyagi Prefectural Medical Institution, it became the Imperial Medical University in 1915 in line with the establishment of Tohoku Imperial University. It has developed into a top-level medical educational, research and practice organization, and its customs and practices have been handed down from generation to generation. Its alumni association, called "Gonryo," includes more than 17,000 members who carry-out excellent work all over the world.

The Postgraduate School of Medicine provides various professors and staffs with superior research abilities that cover most of the medical field, and it admits more than 200 students including master’s and doctoral courses. It produces a high amount of epoch-making research findings to the world.

For more information about Tohoku University Graduate School of Medicine, visit

1 H. Tanaka, S.W. Chiu, T. Watanabe, S. Kaoku and T. Yamaguchi. “Computer-aided diagnosis system for breast ultrasound images using deep learning,” Physics in Medicine & Biology, Vol. 64, No. 23. Dec. 5, 2019.


About SAS

SAS is the leader in analytics. Through innovative software and services, SAS empowers and inspires customers around the world to transform data into intelligence. SAS gives you THE POWER TO KNOW®.

Editorial contacts: