Deep learning radiomics of elastography for diagnosing compensated advanced chronic liver disease: an international multicenter study

Xue Lu, Haoyan Zhang, Hidekatsu Kuroda, Matteo Garcovich, Victor de Ledinghen, Ivica Grgurevic, Runze Linghu, Hong Ding, Jiandong Chang, Min Wu, Cheng Feng, Xinping Ren, Changzhu Liu, Tao Song, Fankun Meng, Yao Zhang, Ye Fang, Sumei Ma, Jinfen Wang, Xiaolong Qi, Jie Tian, Xin Yang, Jie Ren, Ping Liang, Kun WangVis. Comput. Ind. Biomed. Art 2025 · DOI

TL;DR

This study introduces DLRE-X, a novel deep learning radiomics model that integrates both elastography images and liver stiffness measurements to diagnose compensated advanced chronic liver disease (cACLD). The model demonstrates superior accuracy and robustness compared to traditional 2D-SWE methods across diverse international patient populations, etiologies, and ultrasound devices, addressing a critical need for reliable non-invasive diagnostic tools.

What Problem Does This Paper Address?

Accurately diagnosing compensated advanced chronic liver disease (cACLD) is crucial for effective patient management, but current non-invasive methods face significant challenges. Two-dimensional shear wave elastography (2D-SWE) is a promising technique, but its diagnostic accuracy is hampered by inconsistent cut-off values that vary depending on the ultrasound device manufacturer, patient's liver disease etiology, and geographical region. This variability necessitates repeated, complex validation studies for each new system or context, making widespread, reliable application difficult. The paper aims to overcome these limitations by developing a generalized deep learning radiomics model that can perform accurately and consistently across these diverse factors.

Methods

The core of this research is the development of the Generalized Deep Learning Radiomics of Elastography (DLRE-X) model. This model is designed to be robust against variations introduced by different countries, etiologies, and ultrasound device manufacturers.

Key methodological aspects include:

  • Dual-Input Architecture: DLRE-X integrates two types of data:
  • Pseudo-color 2D-SWE images: A region of interest (ROI) is semi-automatically cropped from the elastography image by a radiologist.
  • Textual Liver Stiffness Measurement (LSM) values: The numerical LSM data is also fed into the model.
  • ConvNeXt Architecture: The model utilizes the ConvNeXt architecture, a modern convolutional neural network known for its high performance and efficiency, for feature extraction from the 2D-SWE ROIs.
  • Transfer Learning: Pre-trained weights from the ImageNet dataset are employed to initialize the model, followed by fine-tuning on the study's training data.
  • Data Augmentation: Techniques like random cropping, rotation, flipping, and affine transformation are used to improve model robustness and mitigate overfitting.
  • Dimension Matching: A specific mechanism is implemented to match the dimensions of the LSM and 2D-SWE ROI inputs before feature fusion.
  • Comparison with 2D-SWE: The performance of DLRE-X is directly compared against the standard 2D-SWE cut-off method.

Datasets

The study utilized a comprehensive international, multicenter dataset comprising:

  • 1937 adult patients with chronic liver disease (CLD).
  • Three etiologies: Chronic hepatitis B (CHB), chronic hepatitis C (CHC), and metabolic dysfunction-associated steatotic liver disease (MASLD).
  • Three manufacturers' devices: SuperSonic Imagine (SSI), General Electric (GE), and Mindray.
  • 17 academic medical centers across China, Japan, and Europe.
  • A total of 9472 2D-SWE images.

The dataset was split into:

  • Training set: 1233 patients (6150 images)
  • Internal test set: 309 patients (1543 images)
  • External test set: 395 patients (1779 images)

The European patient data was exclusively used for the independent external test set to validate the model's generalizability without prior exposure during training.

Key Insights and Ideas

  • The Power of Fusion: Combining numerical LSM data with visual elastography image features (radiomics) in a deep learning framework significantly enhances diagnostic performance and robustness. This suggests that neither modality alone captures the full picture of cACLD.
  • Breaking Down Geographic and Device Barriers: DLRE-X demonstrates remarkable consistency in diagnostic accuracy across different countries, etiologies, and ultrasound manufacturers, a critical limitation of the traditional 2D-SWE cut-off method. This suggests a path towards a more universal diagnostic tool.
  • DLRE-X as a "Self-Evolving" System: Unlike static 2D-SWE cut-offs, the deep learning approach has the potential to improve over time as more data is incorporated into the training set, offering a dynamic and adaptive diagnostic solution.
  • Addressing the "One-Size-Fits-All" Problem: The study highlights that a single LSM cut-off is insufficient for cACLD diagnosis globally. DLRE-X offers a data-driven solution that adapts to these variations.

Key Results

  • Superior Overall Performance: In the external test set, DLRE-X achieved a significantly higher Area Under the Receiver Operating Characteristic Curve (AUC) of 0.89 (95%CI: 0.86–0.91) compared to 2D-SWE's AUC of 0.83 (95%CI: 0.79–0.87) (P=0.025).
  • Reduced False Positives/Negatives: The model demonstrated fewer false negatives (49 vs. 114) and false positives (25 vs. 90) compared to 2D-SWE in the external test set.
  • Robustness Across Subanalyses:
    • Countries/Regions: DLRE-X showed consistent AUCs across China, Japan, and Europe (P=0.21), while 2D-SWE's AUC varied significantly (P<0.001).
  • Etiologies: DLRE-X maintained consistent AUCs across CHB, CHC, and MASLD (P=0.27), whereas 2D-SWE showed significant differences (P=0.005).
  • Manufacturers: Both methods showed comparable robustness across SSI, GE, and Mindray devices (DLRE-X P=0.55, 2D-SWE P=0.24).
  • Dual-Input Advantage: The dual-input DLRE-X model outperformed the single-input (ROI only) model in the external test set (AUC 0.89 vs. 0.83, P=0.02) and specifically in the Chinese and CHB subanalyses.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Study Design: The study was retrospective, which can introduce biases.
  • Prevalence Biases: Differences in disease prevalence (e.g., more CHB in China, more MASLD in Europe) and potentially confounding factors like ALT levels could influence results.
  • Sample Size: The sample size for European patients was relatively small, and the study only included three regions, three manufacturers, and three etiologies.
  • Manual ROI Selection: While semi-automatic, the ROI selection still requires radiologist input, which could introduce variability.

Future work should focus on:

  • Larger, Prospective Studies: Including more diverse patient populations, etiologies, and manufacturers.
  • Further Model Improvement: Incorporating more data to enhance the self-evolutionary capability of the DLRE-X model.
  • Exploring Other Imaging Modalities: Investigating the application of similar radiomics approaches to other imaging techniques for liver disease assessment.
  • Clinical Integration: Developing streamlined workflows for the practical application of DLRE-X in clinical settings.

Summary updated

BibTeX
@article{lu2025deep,
  author = {Xue Lu and Haoyan Zhang and Hidekatsu Kuroda and Matteo Garcovich and Victor de Ledinghen and Ivica Grgurevic and Runze Linghu and Hong Ding and Jiandong Chang and Min Wu and Cheng Feng and Xinping Ren and Changzhu Liu and Tao Song and Fankun Meng and Yao Zhang and Ye Fang and Sumei Ma and Jinfen Wang and Xiaolong Qi and Jie Tian and Xin Yang and Jie Ren and Ping Liang and Kun Wang},
  title = {{Deep learning radiomics of elastography for diagnosing compensated advanced chronic liver disease: an international multicenter study}},
  journal = {Vis. Comput. Ind. Biomed. Art},
  year = {2025},
  doi = {10.1186/s42492-025-00199-6},
}
Figure