Fusion of deep and handcrafted features from multi-modal ultrasound data for multi-task liver disease diagnosis.
Authors
Affiliations (7)
Affiliations (7)
- School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, 100084, China.
- Institute for Precision Medicine, Tsinghua University, Beijing, China.
- Department of Ultrasound, Beijing Ditan Hospital, Capital Medical University, Beijing, 100015, China.
- Beijing Engineering Research Center of Mixed Reality and Advanced Display, School of Optics and Electronics, Beijing Institute of Technology, Beijing, China.
- Department of Ultrasound, Beijing Ditan Hospital, Capital Medical University, Beijing, 100015, China. [email protected].
- School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, 100084, China. [email protected].
- Institute for Precision Medicine, Tsinghua University, Beijing, China. [email protected].
Abstract
Chronic liver diseases, including fibrosis, inflammation, and steatosis, often coexist and challenge accurate diagnosis. Current approaches either carry procedural risks or fail to provide comprehensive early assessment. This study aims to develop a non-invasive, multimodal ultrasound-based framework for simultaneous grading of all three conditions. Our method integrates deep semantic features from deep learning networks with handcrafted features from quantitative ultrasound techniques, capturing complementary information from B-mode images, shear wave elastography, and radio-frequency data. Deep features can reveal discriminable and robust high-dimension non-linear relationships, while handcrafted features can capture ultrasonic characteristics related to the three diseases. These features are fused via a lightweight neural network to predict the probability of different grades of the three diseases, with liver biopsy results serving as the reference standard. The proposed method is evaluated on a cohort of 142 patients with chronic hepatitis B. The proposed method achieves AUCs of 0.86, 0.85, 0.89, and 0.93 for fibrosis staging of ≥ F1, ≥ F2, ≥ F3, and F4, respectively; 0.87 and 0.95 for inflammation grading of ≥ A2 and A3, respectively; and 0.83 and 0.93 for steatosis grading of ≥ S1 and ≥ S2, respectively. The proposed method outperforms existing methods and demonstrates robust performance in both early and advanced disease stages. This multi-task framework has potential as a clinically actionable tool for comprehensive liver disease assessment and decision support.