Article: A novel DenseNet-based architecture for liver and liver tumour segmentation Journal: International Journal of Computational Science and Engineering (IJCSE) 2024 Vol.27 No.5 pp.557 - 569 Abstract: The segmentation of the liver and lesions plays an important role in the clinical interpretation and therapeutic planning of hepatic disorders. Analysing volumetric computed tomography scans physically is time-consuming and imprecise. In this work, we propose an automatic segmentation technique based on DenseNet, in which each layer receives feature maps from all layers before it and uses summation instead of concatenation for combining the features. The architecture uses the U-Net model as its basis, and all blocks in each tier of the U-Net architecture are replaced with DenseNet blocks. Utilising DenseNet improves the gradient flow throughout the network, allowing each layer to recognise more diverse and low-complexity features. Other U-Net-based hybrid models use two-dimensional filters that might overlook contextual information. In contrast, the proposed model accepts as input a stack of five adjacent slices and returns a segmentation map for the middle slice. Consequently, this method is highly parameter-efficient and enables enhanced feature extraction. On the LiTs dataset, the experimental findings reveal a liver segmentation accuracy of 94.9% dice score and a tumour segmentation accuracy of 79.2% dice score. On the LiTs competition website, the result indicated that the accuracy was comparable to that of the most effective techniques. Inderscience Publishers - linking academia, business and industry through research

Title: A novel DenseNet-based architecture for liver and liver tumour segmentation

Authors: D.J. Deepak; B.S. Sunil Kumar

Addresses: Department of Information Science and Engineering, RV Institute of Technology and Management, Visvesvaraya Technological University, Belagavi, Bengaluru, India ' Department of Information Science and Engineering, GM Institute of Technology, Visvesvaraya Technological University, Belagavi, Davangere, India

Abstract: The segmentation of the liver and lesions plays an important role in the clinical interpretation and therapeutic planning of hepatic disorders. Analysing volumetric computed tomography scans physically is time-consuming and imprecise. In this work, we propose an automatic segmentation technique based on DenseNet, in which each layer receives feature maps from all layers before it and uses summation instead of concatenation for combining the features. The architecture uses the U-Net model as its basis, and all blocks in each tier of the U-Net architecture are replaced with DenseNet blocks. Utilising DenseNet improves the gradient flow throughout the network, allowing each layer to recognise more diverse and low-complexity features. Other U-Net-based hybrid models use two-dimensional filters that might overlook contextual information. In contrast, the proposed model accepts as input a stack of five adjacent slices and returns a segmentation map for the middle slice. Consequently, this method is highly parameter-efficient and enables enhanced feature extraction. On the LiTs dataset, the experimental findings reveal a liver segmentation accuracy of 94.9% dice score and a tumour segmentation accuracy of 79.2% dice score. On the LiTs competition website, the result indicated that the accuracy was comparable to that of the most effective techniques.

Keywords: liver lesion segmentation; U-Net; DenseNet; computed tomography images; convolutional neural networks; Kaiming initialisation; leaky Relu activation function.

DOI: 10.1504/IJCSE.2024.141340

International Journal of Computational Science and Engineering, 2024 Vol.27 No.5, pp.557 - 569

Received: 24 Mar 2023
Accepted: 01 Sep 2023
Published online: 09 Sep 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: A novel DenseNet-based architecture for liver and liver tumour segmentation

Keep up-to-date