The early detection of liver cancer is crucial in increasing the patient's survival rate. However, detecting malignant liver tumors in the early stage using non-intrusive methods such as computed tomography (CT) imaging has proven to be a challenging task. Thus, the automatic segmentation of liver tumors using deep learning can significantly impact this situation. This study proposes a transfer learning ResNet50-UNET (RU-Net) and EfficientNet hybrid system to classify and segment liver tumors from abdominal CT images in the LiTS database. The model employs the lightweight yet powerful EfficientNetB0 to detect liver tumors and the combined ability of ResNet50 and UNET to perform deep feature extractions and precise segmentations while training on limited data. First, liver segmentation was performed to narrow the region of interest (ROI) using a RU-Net model. The images were then classified as tumorous and non-tumorous with EfficientNetV2B0. Finally, tumor segmentation was done on the tumorous CT images using the same RU-Net architecture. The liver segmentation, tumor classification, and tumor segmentation results achieved a dice similarity coefficient (DSC) of 98.2%, an accuracy of 96.7%, and a DSC of 94.6%, respectively.