A Novel Method for Detecting Liver Tumors - combining Machine Learning with Medical Imaging in CT Scans using ResUNet

Q: Why is liver tumor segmentation difficult in abdominal CT images?

Segmentation is hindered by overlapping intensity values and the unpredictable locations, shapes, and appearances of soft tissues, which reduce boundary accuracy.

Q: Which models are compared, and how do their accuracies differ?

The study reports 2D CNN training accuracy of 96.47% and auto-encoder accuracy of 95.63%, while ResUNet shows the least accurate results among the compared approaches and KNN achieves 86% accuracy; MLP reaches about 28%.

Q: What performance indicators show the proposed method’s effectiveness?

The method’s ROC performance ranges from 0.99 to 1 for both networks, and statistical testing indicates a significant difference with p-value < 0.05 versus several existing algorithms.

Aug 4, 2026

Upload by Theodore

Machine learning optimization methods play a key role in identifying liver tumors, with MRI, CT, and ultrasonography used to separate tumor from liver tissue. Accurate segmentation in computed abdominal CT images remains challenging due to overlapping intensity values and the unpredictable positions, shapes, and appearance of soft tissues. The proposed technique improves overall tumor identification accuracy versus prior state-of-the-art models. Reported results show 2D CNN training accuracy of 96.47%, an auto-encoder accuracy of 95.63%, and strong recall for 2D CNN at 95%. ROC regions range from 0.99 to 1, with statistical significance (p < 0.05) compared to multiple algorithms.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Abstract

Keywords

Introduction

2024 International Conference on Integrated Circuits and Communication Systems (ICICACS) | 979-8-3503-1755-8/24/$31.00 ©2024 IEEE | DOI: 10. 1 109/ICICACS60521.2024. 10499001 2024 International Conference on Integrated Circuits and Communication Systems (ICICACS) . A Novel Method for Detecting Liver Tumors combining Machine Learning with Medical Imaging in CT Scans using ResUNet 1st P. Manjula Department of Logicnet Saveetha School of Engineering, Saveetha Institute of Technical and Medical Sciences Saveetha University Chennai, Tamilnadu, India [manjulasai.sse@saveetha.com](manjulasai.sse@saveetha.com) 2nd K. Krishnakumar Department of Quantum Intelligence Saveetha School of Engineering, Saveetha Institute of Technical and Medical Sciences Saveetha University Chennai, Tamilnadu, India krishnakumarkathiresan.sse@saveetha. com 3rd Smitha GL Department of Quantum Intelligence Saveetha School of Engineering, Saveetha Institute of Technical and Medical Sciences Saveetha University Chennai, Tamilnadu, India [smithagl.sse@saveetha.com](smithagl.sse@saveetha.com) 4th S. Pandiaraj Department of Quantum Intelligence Saveetha School of Engineering Saveetha Institute of Technical and Medical Sciences Saveetha University Chennai, Tamilnadu, India [pandiarajs.sse@saveetha.com](pandiarajs.sse@saveetha.com) 5th M. Prakash Department of Quantum Intelligence Saveetha School of Engineering Saveetha Institute of Technical and Medical Sciences Saveetha University Chennai, Tamilnadu, India [prakashm.sse@saveetha.com](prakashm.sse@saveetha.com) Abstract—The utilization of machine learning optimization methods is of utmost importance in the identification of liver tumors, attracting considerable interest in this domain. After obtaining a liver tissue sample, magnetic resonance imaging (MRI), computed tomography (CT), and ultrasonography (US) are used as imaging techniques to separate the tumor and liver. Nevertheless, the utilization of shades of gray and forms is insufficient for achieving accurate segmentation in computed abdominal CT images, mostly because of the presence of overlapping intensities and the unpredictable placements and shapes of soft tissues. The results demonstrate that our proposed technique outperforms previous state-of-the-art models in terms of overall accuracy in tumor identification. The 2D Convolutional Neural Network (CNN) model had a remarkable training accuracy of 96.47%, while the auto-encoder network closely followed with an accuracy of 95.63%. In addition, the 2D CNN network exhibited an impressive average recall rate of 95%, beating the auto-encoder network's rate of 94%. The ROC curve regions for both networks exhibited remarkable performance, with values ranging from 0.99 to 1. Out of the many machine learning approaches used, the ResUNet had the least accurate results, while the K-Nearest Neighbors (KNN) achieved the greatest accuracy rate of 86%. Conversely, the MLP demonstrated a paltry accuracy rate of about 28%. The statistical tests conducted in this study indicated a significant difference (p-value < 0.05) between the suggested approach and many existing machine learning algorithms. Keywords—Brain Tumor, MRI, Deep Learning, Artificial Intelligence, Transfer Learning, ResUNet I. INTRODUCTION The field of machine learning has shown significant growth and has attracted significant interest from both academics and professionals. It has become a prominent and significant topic of research, with enormous importance in numerous disciplines such as machine translation, speech recognition, picture recognition, and recommendation systems. Optimization is a crucial component of machine learning. The majority of machine learning algorithms include creating an optimization model and obtaining the parameters in the objective function using the given data. The prevalence of numerical optimization methods in the era of abundant data has a substantial influence on the extensive adoption

FAQs

Why is liver tumor segmentation difficult in abdominal CT images?

Segmentation is hindered by overlapping intensity values and the unpredictable locations, shapes, and appearances of soft tissues, which reduce boundary accuracy.

Which models are compared, and how do their accuracies differ?

The study reports 2D CNN training accuracy of 96.47% and auto-encoder accuracy of 95.63%, while ResUNet shows the least accurate results among the compared approaches and KNN achieves 86% accuracy; MLP reaches about 28%.

What performance indicators show the proposed method’s effectiveness?

The method’s ROC performance ranges from 0.99 to 1 for both networks, and statistical testing indicates a significant difference with p-value < 0.05 versus several existing algorithms.

Aug 4, 2026

Upload by Theodore