Noninvasive Staging of Hepatic Fibrosis in Patients with Autoimmune Liver Disease Using Deep Learning.
Authors
Affiliations (5)
Affiliations (5)
- Shandong First Medical University (Shandong Academy of Medical Sciences), Jinan, Shandong 250117, China (H.Y.); Ultrasound Medicine Department, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong 250021, China (H.Y., Y.G.).
- School of Radiology, Shandong First Medical University and Shandong Academy of Medical Sciences, Taian, China (S.Z., Y.P., J.Q.).
- Departments of Ultrasound Medicine Beijing Chaoyang Hospital, Capital Medical University, South Road 8, Gongti, Chaoyang, Beijing 100020, China (Q.S.).
- School of Radiology, Shandong First Medical University and Shandong Academy of Medical Sciences, Taian, China (S.Z., Y.P., J.Q.); Department of Radiology, The First Affiliated Hospital of Shandong First Medical University, Jinan, Shandong 250000, China (J.Q.).
- Ultrasound Medicine Department, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong 250021, China (H.Y., Y.G.). Electronic address: [email protected].
Abstract
Accurate staging of hepatic fibrosis is essential for guiding immunosuppressive and antifibrotic therapies. However, percutaneous liver biopsy, the current reference standard, remains invasive and is subject to sampling errors and interobserver variability. To address these limitations, we developed and validated a noninvasive deep learning model using routine two-dimensional B-mode ultrasound for fibrosis staging in patients with autoimmune liver disease (AILD). We retrospectively enrolled 245 consecutive patients with AILD and randomly assigned them to the training set (60%), validation set (20%), and internal testing set (20%). Additionally, 61 biopsy-confirmed patients with AILD from another hospital were recruited as an external testing set. A deep learning model was constructed using the ResNet34 network architecture based on two-dimensional B-mode ultrasound images to evaluate its performance in liver fibrosis staging. Model performance was assessed using metrics such as macro- and microarea under the curve (AUC). Calibration curves and decision curves were employed to evaluate model goodness-of-fit and clinical utility, and class activation mapping was used for model interpretation. The model demonstrated robust performance across different datasets. In the internal and external test sets, the macroaverage AUCs were 0.812 (0.692-0.901) and 0.801 (0.688-0.902), respectively, while the microaverage AUCs were 0.819 (0.717-0.900) and 0.847 (0.761-0.911), respectively. The calibration and decision curves indicated favorable goodness-of-fit and clinical utility, and the class activation maps revealed the model's decision-making rationale, enhancing interpretability. The model demonstrated robust diagnostic performance for the noninvasive staging of hepatic fibrosis in patients with AILD.