There is no doubt that breast cancer has the highest prevalence among female patients. In addition, it is the only one among the six major cancers that has shown an increasing trend over the past 20 years.
Early detection and treatment of breast cancer increases the survival rate. However, after stage 3, his survival rate drops significantly to less than 75%. In other words, early detection through regular health checkups is important to reduce patient mortality. Recently, POSTECH’s research team developed an AI network system for ultrasound examinations to accurately detect and diagnose breast cancer.
A team of researchers at POSTECH, led by Prof. Chulhong Kim (Department of Convergence IT Engineering, Electrical Engineering, Mechanical Engineering) and Sampa Misra and Chiho Yoon (Department of Electrical Engineering) developed deep learning. Based multimodal fusion network for breast cancer segmentation and classification using B-mode and strain elastography ultrasound images. The results of this study were published in Bioengineering & Translational Medicine.
Ultrasonography is one of the important medical imaging modalities for evaluating breast lesions. To distinguish between benign and malignant lesions, computer-aided diagnosis (CAD) systems have provided radiologists with tremendous assistance by automatically segmenting and identifying lesion features.
Here the team presented a deep learning (DL)-based method to segment lesions and classify them as benign or malignant using both B-mode and strain elastography (SE-mode) images. First, the team utilized the weighted skip connection method to build a ‘weighted multimodal U-Net (W-MM-U-Net) model’ that segmented lesions by assigning optimal weights to different imaging modalities. bottom. We also proposed a “multimodal fusion framework (MFF)” for cropped B-mode and SE-mode ultrasound (US) lesion images to classify benign and malignant lesions.
MFF consists of an integrated function network (IFN) and a decision network (DN). Unlike other recent fusion methods, the proposed MFF method can simultaneously learn complementary information from a convolutional neural network (CNN) trained on his US images in B-mode and SE-mode. The features of CNN are ensembled using a multimodal EmbraceNet model, while DN uses those features to classify images.
The method predicts 7 benign patients to be benign in 3 out of 5 trials and 6 malignant patients to be malignant in 5 out of 5 trials, according to experimental results on clinical data. predicted to be. This means that the proposed method is superior to conventional single- and multimodal methods and could potentially improve radiologists’ classification accuracy for breast cancer detection in US images. I have.
Professor Chulhong Kim explains: “We trained each deep learning model and an ensemble model simultaneously to achieve much better classification performance than traditional single-modal or other multimodal methods.”
Original: AI-powered ultrasound images to detect breast cancer
Than: Pohang University of Science and Technology
