Summary
With the advancement of technology, besides the 2-D handheld US (HHUS), the 3-D automated whole breast ultrasound (ABUS) has been developed for screening the whole breast clinically. Hence, the aim of this project is to develop and deploy a computer-aided detection (CADe) and diagnosis (CADx) system on cloud system with the deep learning and cloud technology for the tumor detection, segmentation, and diagnosis automatically. Up to now, the 3-D ABUS CADe and CADx and the 2-D HHUS CADe systems have been developed for tumor detection in ABUS and diagnosis in HHUS image. In 3-D ABUS CADe, while sensitivity is at 95%, the corresponding average FP/per pass is 2.6. In 3-D ABUS CADx, the average dice coefficient of tumor segmentation is 0.88 and the accuracy, sensitivity, and specificity of tumor diagnosis are 84.9%, 87.2%, and 82.6%. In 2-D HHUS CADe systems, the detection rate and the dice coefficient are 0.91 and 0.85. Furthermore, the accuracy, sensitivity, specificity, and AUC of lymph node metastasis are 81.1%, 81.4%, 80.9%, and 0.8054, respectively.
 
Applications
We have developed two applications. The first one is an improving CADe and an additional CADx system applied on 3-D ABUS image for tumor detection and classification. The second one is a CADe system employed on 2-D HHUS for tumor detection.
 
Advantages
In 3-D ABUS image, it can be reviewed and the suspicious lesion will be detected automatically about 0.8 seconds per pass with high sensitivity and low FP by the CADe system. Then, the additional CADx system will provide information about tumor such as malignant or benign. In 2-D HHUS CADe systems, the detection method can find out the tumor location, separate the tumor region from the original image, and predict the lymph node status.
 
Keywords
Breast cancer, automated whole breast ultrasound, computer-aided tumor detection, computer-aided diagnosis, deep learning, cloud system