Development of a hybrid deep learning-based framework for liver fibrosis classification using ultrasound images.

Background and aims: Liver fibrosis is a progressive accumulation of extracellular matrix proteins with distortion of hepatic architecture and can progress to cirrhosis or hepatocellular carcinoma. Biopsy remains the diagnostic gold standard, however, its invasive nature, sampling error, and cost limit routine use. Ultrasound imaging provides a safer, more accessible option but depends on operator expertise and subjective interpretation. Existing deep learning approaches for fibrosis assessment often rely on small datasets or perform only binary classification. This study aimed to develop a hybrid deep learning model combining ResNet50 and VGG16 for automated multi-class classification (F0-F4), enhancing diagnostic accuracy, reducing biopsy reliance, and supporting affordable, interpretable liver disease assessment. Methods: The total of 6323 ultrasound image samples with METAVIR system labels ranging from F0 to F4 was downloaded from Kaggle. After data preprocessing, 80:20 splits were made for training and testing. A hybrid model consisting of fine-tuned ResNet50 and VGG16 was used for classification of fibrosis stages. Model performance was statistically evaluated using sensitivity, specificity, and area under the ROC curve (AUC) for each fibrosis stage, averaging across classes to address imbalance. Robustness and reproducibility were assessed by calculating 95% confidence intervals (CI) for all performance metrics through bootstrap resampling. Grad-CAM was used to interpret the model's predictions. Results: The hybrid model was also successful in achieving the highest peak in testing accuracy, reaching 86.64%, compared to the other models (55.26% for ResNet50, 72.73% for VGG16). The classification was also high for the hybrid model, with the highest values for the macro AUC and weighted AUC at 96.79% and 97.79%, respectively. The highest predicted probabilities were seen for the F0 and F4 stages, which were correctly classified, with the Grad-CAM heatmaps showing high focus on the fibrotic regions. Conclusion: The hybrid model achieved good diagnostic results with high sensitivity, specificity, and confidence. The Grad-CAM images validated that the model was focusing on significant areas, as shown by the heat map, which proves that it has potential as a non-invasive, accurate, and interpretable tool for automated liver fibrosis staging using ultrasound images.