Preliminary Exploration of Discriminative Correlation Filter Network for Real-time Tracking of Inconspicuous Focal Liver Lesions on Conventional Ultrasound.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Ultrasonics, The Eighth Affiliated Hospital of Sun Yat-sen University, Shenzhen, Guangdong Province, China; Institute of Biomedical Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China.
- Department of Medical Ultrasonics, The Eighth Affiliated Hospital of Sun Yat-sen University, Shenzhen, Guangdong Province, China. Electronic address: [email protected].
- Department of Medical Ultrasonics, The Eighth Affiliated Hospital of Sun Yat-sen University, Shenzhen, Guangdong Province, China.
- Institute of Biomedical Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China.
- Institute of Biomedical Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China. Electronic address: [email protected].
Abstract
This study aimed to explore the feasibility of a discriminative correlation filter network (DCFNet)-based algorithm for positioning inconspicuous focal liver lesions (FLLs) on conventional US by tracking surrounding anatomical landmarks (SALs). We established an algorithm that exploits the relative position characteristics of SALs to achieve tracking of inconspicuous FLLs in real-time US. Multiple SALs were tracked simultaneously using DCFNet, a fast deep learning-based tracking framework. The proposed DCFNet-based algorithm could assist real-time FLLs tracking on US and further guide puncture or ablation. We designed US-visible and US-invisible liver tumor phantoms for tracking and puncture performance evaluation. Algorithm performance was evaluated via matching accuracy, precision and positional error. Puncture verification was assessed in terms of success rate and guidance time on US-invisible liver tumor phantoms. A total of 20 US-visible phantoms, 20 US-invisible phantoms, and initial clinical cases were included. the algorithm achieved a tracking accuracy of 93.2%, precision of 92.9% and mean tumor localization error of 0.7 ± 0.4 mm in US-visible phantoms. For US-invisible phantoms, the puncture success rate was 94.3% with an average duration of 4 (3-5) s. The DCFNet-based algorithm could be used as a potential approach for locating and tracking inconspicuous FLLs on US, with preliminary clinical feasibility that warrants further large-scale validation.