A Novel Deep Learning Framework for Liver Fibrosis Staging and Etiology Diagnosis Using Integrated Liver–Spleen Elastography

Objectives: Liver fibrosis staging and etiology diagnosis are critical for patient management, but non-invasive methods remain challenging. This study aims to evaluate the performance of radiomics models using 2D shear wave elastography (2D-SWE) of the liver and spleen for liver fibrosis staging and etiology differentiation, comparing them with serum biomarkers and conventional ultrasound. Methods: A retrospective analysis was conducted on 198 patients with liver fibrosis confirmed by biopsy. Radiomics features were extracted from the liver and spleen grayscale and 2D-SWE images. Machine learning (ML) and transfer learning (TL) models were established for fibrosis staging and etiology diagnosis. Model performance was evaluated according to receiver operating characteristic (ROC) curves. Results: For fibrosis staging, 2D-SWE-based models outperformed grayscale and serum biomarkers. The combined liver–spleen TL model achieved exceptional validation performance (AUCs 0.99 for S4, 0.98 for ≥S3, 1.00 for ≥S2). For etiology diagnosis, the liver 2D-SWE TL model and the combined liver–spleen TL model achieved AUCs of 0.97 and 0.94, respectively, significantly outperforming ML models in terms of AUC. Conclusions: Integrating liver and spleen 2D-SWE radiomics with TL significantly improves non-invasive liver fibrosis staging and etiology diagnosis, offering superior accuracy over conventional methods. This approach holds promise for clinical application, though further validation is needed.