Abstract
The staging of liver fibrosis is essential in influencing the final clinical management and treatment intervention that depends on proper and prompt diagnosis. The suggested multi-stream deep learning architecture that also involves the combination of EfficientNetB0 and ResNet50 models may be regarded as a feature-level fusion solution that utilizes the most sophisticated normalization/regularization techniques. A full fusion model was also implemented, reaching a classification accuracy of 99.45% and a loss of 0.0295, which was higher than in any of the individual models (EfficientNetB0 with 98.50% and ResNet50 with 99.13%). In order to establish the robustness of the model, ablation analysis was carried out in detail, and it investigated the impact of an architectural element, including batch normalization, dropout layers, and fusion strategies. The results support the advantage of both normalization and dropout in terms of better generalization ability, but the feature fusion dramatically outperforms a simple concatenation in respect to discriminative ability. The findings reveal the strength of the multi-stream approach offered in the problem of homeless liver fibrosis stage classification, which means that it can be used in clinical practice.