Diagnosis of early hepatic fibrosis is difficult since currently available non-invasive approach is hardly sensitive enough to identify early hepatic fibrosis in the course of treatment. We suggest a convolutional neural network as a deeper learning model to identify MRI-based fibrosis in the liver in order to classify it. As it merges both DenseNet121 and ResNet50 and EfficientNet-B0 into a single fusion system, the model is better than diagnosing with the traditional methods. Transfer learning and custom weighted feature fusion is implemented, and the training was proven with a stratified train-test split. The performance of classification is measured by ROC curves and class-specific metrics, thus allowing specific evaluation with regard to different fibrosis stages. DICOM files of medical image data are pre-processed whereas middle-slice selection represents patient series. Results indicated that the fusion model could conduct better detection than individual models on a variety of fibrosis stages as demonstrated by micro-average AUC-ROC. Such an approach could result in more precise but non-invasive diagnosis and better clinical outcomes of patients with chronic liver disease.