Deep learning technique for automatic liver and liver tumor segmentation in CT images

Abstract

Segmenting the liver and tumors from computed tomography (CT) scans is crucial for medical studies utilizing machine and deep learning techniques. Semantic segmentation, a critical step in this process, is accomplished effectively using fully convolutional neural networks (CNNs). Most Popular networks like UNet and ResUNet leverage diverse resolution features through meticulous planning of convolutional layers and skip connections. This study introduces an automated system employing different convolutional layers that automatically extract features and preserve the spatial information of each feature. In this study, we employed both UNet and a modified Residual UNet on the 3Dircadb (3D Image Reconstruction for computer Assisted Diagnosis database) dataset to segment the liver and tumor. The ResUNet model achieved remarkable results with a Dice Similarity Coefficient of 91.44% for liver segmentation and 75.84% for tumor segmentation on 128 × 128 pixel images. These findings validate the effectiveness of the developed models. Notably both models exhibited excellent performance in tumor segmentation. The primary goal of this paper is to utilize deep learning algorithms for liver and tumor segmentation, assessing the model using metrics such as the Dice Similarity Coefficient, accuracy, and precision.