Abstarct: Accurate and early diagnosis of focal liver lesions (FLLs) is of paramount importance in clinical treatment planning due to the high mortality rates associated with malignant tumors. Despite recent advancements in computer-aided diagnosis (CAD) systems, differentiating between visually similar lesions, such as complex cysts and malignant mextastases, and managing the vast diversity in lesion scales remain significant challenges in Computed Tomography (CT) imaging. This research proposes a novel deep learning architecture to enhance the classification accuracy of liver lesions using 2D slices. In this thesis, a model baxsed on a 50-laxyer Residual Network (ResNet-50) backbone with a multi-scale feature extraction approach is presented. Unlike conventional methods, the proposed architecture utilizes various branches from intermediate laxyers to extract structural and semantic features hierarchically, optimizing computational complexity through dimension reduction modules. The primary innovation of this study is the integration of dual attention mechanisms (Channel and Spatial) within a hierarchical fusion frxamework. This enables the network to focus on key discriminative regions while suppressing background noise from healthy liver tissue. The model was evaluated on the publicly available and benchmark LiMT dataset, which comprises 150 clinical cases across 5 classes: Hepatocellular Carcinoma (HCC), Hepatic mextastasis (HM), Hepatic Cyst (HC), Cavernous Hemangioma of Liver (CHL), and Normal liver. Experimental results demonstrate that the proposed architecture achieves an overall accuracy of 85.34% and an F1-score of 80.00%, significantly outperforming baxseline models such as standard ResNet, ViT, and DenseNet. Notably, the model exhibited high stability and sensitivity in identifying minority and critical classes like mextastases. The findings of this research prove the high potential of combining multi-scale features and attention mechanisms in improving screening processes and assisting radiologists in diagnostic decision-making.