Modified U-Net for liver cancer segmentation from computed tomography images with a new class balancing method

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver cancer is a major global health concern, often diagnosed via computed tomography (CT).
• Deep learning techniques are increasingly utilized for segmenting organs and tumors in medical scans.
• Accurate segmentation is crucial for timely liver cancer diagnosis and treatment planning.

Purpose of the Study:

• To develop and evaluate a deep learning model for segmenting liver and liver tumors from abdominal CT scans.
• To improve the efficiency and accuracy of liver cancer diagnosis by automating segmentation.
• To modify the UNet architecture for enhanced performance in liver and tumor segmentation.

Main Methods:

• The study adapted the UNet deep learning architecture for medical image segmentation.
• Key modifications included reducing convolutional filters and adding batch normalization and dropout layers.
• A novel class balancing method was incorporated to address data imbalance issues.

Main Results:

• The modified UNet achieved a dice score of 0.96 for liver segmentation and 0.74 for tumor segmentation within the liver.
• Segmentation of tumors from abdominal CT scans yielded a dice score of 0.63.
• The proposed method demonstrated improvements of 0.01 for liver and 0.11 for tumor segmentation compared to existing approaches.

Conclusions:

• The modified UNet architecture offers improved performance for liver and tumor segmentation in CT images.
• Network complexity was reduced, and segmentation accuracy was enhanced through architectural modifications and class balancing.
• The algorithm showed limitations in segmenting small and irregularly shaped tumors, indicating areas for future research.