Automated liver segmentation from CT images using modified ResUNet

Recent research has focused on multimodal medical image segmentation. A cascaded V-net and H-DenseUNet approach have improved Dice scores, but at the expense of high computational complexity. Optimization of feature extraction through detail-enhanced convolution improved performance, though this method remained limited to single-organ segmentation. Another approach introduced deep supervision and dilated convolution, which enhanced accuracy but still fell short recognizing boundary tumors.

Liver lesion detection relies on precise computed tomography (CT) image segmentation. However, traditional methods face challenges due to the liver's variable morphology, overlap with adjacent organs, and anatomical complexity. Semantic segmentation techniques which allow for pixel-wise classification, offer significant analytical potential, while deep learning models like UNet have greatly advanced segmentation performance.

In a study published in Gastroenterology & Endoscopy, a team of researchers in India proposed a new model to address the high computational costs and inadequate adaptability to complex cases of existing models. By integrating residual networks to strengthen feature extraction, their approach was validated on the public dataset 3Dircadb for automated and precise liver disease diagnosis.

“The ResUNet model integrates the UNet encoder-decoder architecture with ResNet residual blocks, consisting of 52 layers, including convolutional, ReLU activation, max pooling, and up-sampling layers. Feature propagation was optimized via skip connections, with final segmentation results output using SoftMax classification,” shares  lead author R.V. Manjunath.

Experiments demonstrate that ResUNet achieved a Dice coefficient of 93.08% and accuracy of 98.57% on the test set, significantly outperforming traditional models, while ablation experiments revealed that the configuration with three residual blocks yields optimal performance.

“The model exhibits stable performance in cross-dataset tests on LiTS and 3Dircadb, with training curves indicating convergence after 250 epochs,” adds Manjunath. “Compared to no data augmentation, the enhanced approach reduces VOE by 3.77%, demonstrating the critical role of data diversity in improving performance.”

Overall, the results validated the effectiveness of the ResUNet model in CT liver segmentation, achieving high-precision results through architectural optimization and data augmentation.

“This technology has the potential to lower the barrier for primary healthcare diagnosis and advance the widespread adoption and intelligent development of early liver disease screening,” says Manjunath. “Going forward, we want to scale to 1024×1024 high-resolution images, multi-organ segmentation, and real-time clinical data integration, while improving training efficiency (currently requiring 3531 minutes),”

###

Contact the author: R.V. Manjunath, Department of Electronics & Communication Engineering, Dayananda Sagar Academy of Technology and Management, Bangalore, 82, Karnataka, India, manjunathrv@dsatm.edu.in

The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 200 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).