Diagnosis of Small Focal Liver Lesions (≤2 cm): A Deep Learning Approach Based on B-Mode Ultrasound and Contrast-Enhanced Ultrasound.
Authors
Affiliations (8)
Affiliations (8)
- Department of Medical Ultrasound, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China.
- Department of Ultrasound, The First Affiliated Hospital of Anhui Medical University, Hefei, Anhui, China.
- Department of Ultrasonic Imaging, Xiangya Hospital, Central South University, Changsha, Hunan, China.
- Department of Ultrasound Medicine, Hunan Provincial People's Hospital, Changsha, Hunan, China.
- Britton Chance Center and MoE Key Laboratory for Biomedical Photonics, Wuhan National Laboratory for Optoelectronics-Huazhong University of Science and Technology, Wuhan, Hubei, China.
- Department of Ultrasound, Seventh Medical Center, Chinese PLA General Hospital, Beijing, Beijing, China.
- Department of General Internal Medicine, Kliniken Hirslanden Beau-Site, Bern, Switzerland.
- Department of Medical Ultrasound, Tongji Hospital, Tongji Medical College and State Key Laboratory for Diagnosis and Treatment of Severe Zoonotic Infectious Diseases, Huazhong University of Science and Technology, Wuhan, Hubei, China. Electronic address: [email protected].
Abstract
To develop and validate a deep learning (DL) model based on feature fusion with B-mode ultrasound (BMUS) and contrast enhanced ultrasound (CEUS) images for non-invasive diagnosis of benign and malignant focal liver lesions (FLLs), especially for small FLLs (≤2.0 cm). A retrospective dataset of 687 patients who were diagnosed with FLLs by BMUS and underwent CEUS between September 2016 and March 2020 was collected in this study. The dataset was divided into a training set (80%) and a test set (20%). For the feature fusion strategy can leverage the complementary information from each feature to obtain the complete information of an image. We innovatively designed a DL model based on the ResNet-18, integrated with the feature fusion, to extract and consolidate deep feature representations from both BMUS images and three key CEUS phases (arterial, portal venous, and delayed). The performance of the model was compared to that of four radiologists. Subgroup analyses were performed to assess model performance in patients with small FLLs (≤2.0 cm), different genders, hepatitis B virus (HBV) infection status, and cirrhosis. The DL model achieved an area under the curve (AUC) of 0.994 in the test set, significantly surpassing four participating radiologists (AUCs ranging from 0.787 to 0.930; all p < 0.05). In terms of sensitivity and specificity, the model achieved 97.53% and 94.74%, respectively, showing no significant difference compared to two of the experienced radiologists (all p > 0.05). Furthermore, in subgroup analyses, including patients with small FLLs (≤2.0 cm), different genders, hepatitis B virus (HBV) infection, and cirrhosis status, the model consistently outperformed all four radiologists, achieving AUCs between 0.990 and 0.999 across subgroups. By harnessing BMUS and CEUS images of patients with FLLs, the feature fusion-based DL model offers an effective diagnostic tool for discerning the benign or malignant nature of FLLs.