Enhancing Liver Tumour Segmentation in CT Images Using Dilated Residual Capsule Networks

• Published in: Journal Europeen des Systemes Automatises Vol. 57; no. 6; p. 1775
• Main Authors: Allenki, Jyoshna, Soni, Hemant Kumar
• Format: Journal Article
• Language: English
• Published: Edmonton International Information and Engineering Technology Association (IIETA) 01.12.2024
• Subjects: Accuracy; Architecture; Automation; Classification; Datasets; Deep learning; Liver; Machine learning; Medical imaging; Neural networks; Semantics; Tumors
• ISSN: 1269-6935; 2116-7087
• DOI: 10.18280/jesa.570625

Liver CT images play a crucial role in the early diagnosis of liver disorders and have proven effective in identifying chronic liver disease, which may lead to fatal outcomes. This imaging technique provides detailed cross-sectional views, allowing for precise detection of abnormalities, aiding in timely intervention, and improving patient prognosis. The detection process of chronic liver disease should be carried out with meticulous accuracy. Due to the inherent complexities involved and the presence of ambiguities in CT images, segmentation approaches have not yet reached the pinnacle of accurate and reliable performance required for clinical application. Recently, the emergence of machine learning and deep learning algorithms has provided valuable insights into achieving a more accurate segmentation process. However, these existing deep learning algorithms suffer from several challenges that hinder segmentation performance. Hence, independent deep learning algorithms require further refinement to handle CT liver images effectively. To address this problem, this research article proposes a fully automated, robust, and accurate segmentation of CT liver images based on a deep neural network architecture that adopts dilated residual networks integrated with powerful capsule networks. This proposed network combines the strengths of capsule networks and ResNet-50 architectures to achieve better segmentation results. Extensive experimentation is conducted using 100 healthy subjects, and 131 contrast-enhanced image data are used for training, while 70 CT images are used for testing. Furthermore, the proposed model is evaluated using performance metrics such as DICE, Intersection over Union (IoU), precision, and recall. To demonstrate the superiority of the suggested network, its segmentation performance is compared with that of existing state-of-the-art deep learning architectures. The results demonstrate that the suggested model achieved 0.98 DICE, 0.95 IoU, 99.2% precision, and 99.1% recall, respectively, surpassing various existing models used for liver CT image segmentation.