Purpose <p>Liver tumors, the sixth most common form of cancer, are highly heterogeneous. Precise automated segmentation of such tumors is crucial for accurate prognosis and treatment planning. Though manual segmentation is reliable, it is time-consuming and prone to human error and variables. In this study, we propose an automated segmentation technique using CNN, which is vital in addressing these issues and supporting precision medicine.</p> Methods <p>This work proposes a dual-stage segmentation process for segmenting liver tumors in abdominal CT scans. A U-Net architecture driven by deep learning was initially implemented to segment the liver. Further, Radiomic features such as first-order statistics, shape, and texture have been extracted from the segmented liver. These features supervise a linearly modulated U-Net (RFiLM-Net) framework to refine tumor segmentation. The proposed method uses advanced convolutional neural networks for feature extraction and incorporates feature-induced linear modulation for enhanced accuracy. Further, A fivefold cross-validation scheme confirms the robustness and generalizability of the proposed model.</p> Results <p>The model achieved a dice similarity coefficient of 0.92 for liver segmentation and 0.87 for tumor segmentation, outperforming existing state-of-the-art. The method (RFiLM-Net) proposed high accuracy, precision, and recall rates of 99%, 98%, and 99%, respectively. Integrating radiomic features into the U-Net architecture significantly improved segmentation performance, which was confirmed through fivefold cross-validation. Our findings indicate that the two-stage RFiLM U-Net model substantially advances automated liver tumor segmentation, which renders it a viable option for immediate clinical application.</p> Conclusion <p>RFiLM-Net significantly improves over existing automated segmentation methods. This demonstrates robust performance and readiness for use in identifying liver tumors in current clinical practice.</p>