Page 66 - AIH-2-2

Artificial Intelligence in Health

ORIGINAL RESEARCH ARTICLE
The development of neural networks with

intra- and inter-block dense connections for
improved liver tumor segmentation

3
1
Yizhuo Xu 1† , Shen Huang * , Shanshan Zhao , Yiyang Zhang , and
2†
Ziyang Huang 1
1 Department of Information Engineering, School of Electronic Information and Communications,
Huazhong University of Science and Technology, Wuhan, China
2 School of Public Health, Tongji Medical College, Huazhong University of Science and Technology,
Wuhan, China
3 Department of Personnel, Kunming Medical University, Kunming, China

Abstract

Automatic liver tumor segmentation is an essential part of computer-aided diagnosis
systems. Despite the significant progress made by fully convolutional neural networks
(FCNs) in recent years, existing methods fail to effectively segment multiscale
tumors and accurately delineate tumor boundaries. This indicates that the single-
† These authors contributed equally
to this work. connection network structure overly relies on the high-level semantic information of
the image but fails to fully utilize the spatial information of the target across different
*Corresponding author: levels. To solve these problems, we introduce a novel end-to-end segmentation
Shen Huang
2
(huangshen721@163.com) network, named intra- and inter-block densely connected FCN (I -DenseFCN). This
network comprising intra-block and inter-block dense connections in the encoder
Citation: Xu Y, Huang S,
Zhao S, Zhang Y, Huang Z. The and decoder, respectively, is developed to effectively handle the segmentation of
development of neural networks multiscale tumors. We also designed a hybrid loss function to achieve more accurate
with intra- and inter-block dense tumor boundary delineation by adaptively adjusting the optimization of the cross-
connections for improved liver
tumor segmentation. Artif Intell entropy and Lovász Softmax loss. In addition, we are the first to propose an intensity-
Health. 2025;2(2):60-72. adaptive image pre-processing method that effectively mitigates the differences
doi: 10.36922/aih.5001 between real-world medical scenarios and model inference for liver tumor
Received: September 30, 2024 segmentation tasks. Our training data, derived from the Liver Tumor Segmentation
Challenge (LiTS) public dataset, have been expanded to 25,755 samples of liver and
Revised: November 5, 2024
their tumors through data augmentation techniques. Experimental evaluations on
Accepted: December 3, 2024 the LiTS and 3DIRCADb databases demonstrated the superiority of the proposed
2
Published online: December 19, I -DenseFCN over classical methods. We believe this research will contribute to the
2024 field of tumor segmentation using AI-based medical image processing methods,
Copyright: © 2024 Author(s). which is a promising application in computer-aided diagnosis systems.
This is an Open-Access article
distributed under the terms of the
Creative Commons Attribution Keywords: Liver tumor segmentation; Deep learning; Multiscale tumors; Computed
License, permitting distribution, tomography imaging; Artificial intelligence; AI diagnostics
and reproduction in any medium,
provided the original work is
properly cited.
Publisher’s Note: AccScience
Publishing remains neutral with 1. Introduction
regard to jurisdictional claims in
published maps and institutional According to Global Cancer Statistics 2018, liver cancer is the sixth most prevalent
affiliations. cancer and the fourth leading cause of cancer-related deaths. Accurate liver tumor
1

Volume 2 Issue 2 (2025) 60 doi: 10.36922/aih.5001