3D multi-view tumor detection in automated whole breast ultrasound using deep convolutional neural network. (15th April 2021)

• Record Type: Journal Article
• Authors: Zhou, Yue; Chen, Houjin; Li, Yanfeng; Wang, Shu; Cheng, Lin; Li, Jupeng
• Is Part Of: Expert systems with applications. Volume 168(2021)
• Publication Date: 2021-04-15
• Subjects: Automated breast ultrasound (ABUS) -- 3D detection -- Multi-view detection -- Deep learning -- Majority voting -- Candidates fusion; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.114410
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: In recent years, automated whole breast ultrasound (ABUS) has drawn attention to breast disease detection and diagnosis applications. However, reviewing ABUS volumes is a time-costing task and some subtle tumors may be missed. In this paper, a 3D multi-view tumor detection method is proposed for ABUS volumes. Firstly, a layer connected feature extraction network is designed for Faster R-CNN. Then, orthogonal multi-view slices are reconstructed and detected using this modified Faster R-CNN to extract 2D candidates. Finally, a 3D multi-view position analysis scheme is designed to fuse 2D detection results and get final 3D bounding boxes. The performance of this proposed method is evaluated on a data set of 158 volumes from 75 patients by 5-fold cross-validation. Experimental results show that our method achieves a sensitivity of 95.06% with 0.57 false positives (FPs) per volume. Compared with existing detection methods, the proposed method is more effective and general. Highlights: A new feature extraction network is designed for breast ultrasound image detection. A 3D position analysis scheme is proposed to fuse 2D results and remove FPs. Quantitative and qualitative analysis are shown for 2D and 3D detection results. The proposed 3D detection method gets the highest sensitivity at 0.95 with 0.57 FPs.