Liver histopathology plays a critical role in diagnosing and evaluating fibrosis, a condition often assessed through ultrasound imaging. This research proposes an ensemble deep learning model combining ResNet50 and VGG16 architectures, enhanced with attention mechanisms, to effectively classify liver histopathology and fibrosis images. The dataset, comprising grayscale ultrasound images from five classes, was preprocessed by converting to RGB, resizing, and stratified splitting into training, validation, and testing subsets. The model architecture integrates global average pooling and dense layers for classification, while attention blocks amplify salient features, enhancing predictive accuracy. Comprehensive training incorporated data augmentation and early stopping to mitigate overfitting. The model achieved a test accuracy of 97%, with consistent precision, recall, and F1-scores across all classes. Performance evaluation through confusion matrices, t-SNE feature visualization, and ROC curves with high AUC values validated its robustness and reliability. This approach demonstrates significant potential for automated medical image classification, advancing diagnostic accuracy and efficiency in clinical applications.