Hybrid DenseNet-UNet Model for Accurate Liver Segmentation in CT images

Abstract

Liver segmentation from CT images is a critical and foundational task in medical image analysis, playing a pivotal role in accurate diagnosis, treatment planning, and patient management, particularly in liver-related diseases. The ability to precisely delineate the liver is essential for tasks ranging from assessing liver volume to planning surgical procedures and targeting radiation therapy. In this work, an advanced adaptation of the U-Net architecture, integrating DenseNet121 as its backbone is used. This combination leverages DenseNet’s dense connections, ensuring efficient gradient flow and feature reuse, enhancing learning capability. Preprocessing steps, including resizing images to 256x256 pixels, histogram equalization, normalization, and binary mask conversion, are applied to ensure data consistency and enhance model performance. Two distinct datasets, 3D-IRCADb-01 and LiTS, are used. The Dice Similarity Coefficient (DSC) is used to evaluate the performance of various models. For dataset 3D-IRCADb-01, remarkable DSC scores are achieved, with the highest reaching 96.5%, and accuracy of 99.5%, indicating the effectiveness of the segmentation models. For dataset LiTS, the models excelled further, achieving DSC scores as high as 98.1% and accuracy of 99.7%. After segmentation, regions of interest (ROIs) are extracted, facilitating subsequent medical analysis and diagnosis. These results demonstrate the robustness and accuracy of the proposed model in liver segmentation tasks.