Liver tumor segmentation and classification using FLAS-UNet++ and an improved DenseNet
Quchen Peng1,2, Yunqi Yan3, Lijun Qian3
1Department of Electronic Engineering, Fudan University, Shanghai, China.
Summary
This study presents an improved deep learning method for accurate liver tumor segmentation and classification from CT images. The approach enhances diagnostic accuracy, offering significant clinical value for liver tumor treatment.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Liver tumors are a significant health concern, with accurate diagnosis crucial for effective treatment.
- Distinguishing benign from malignant liver tumors presents challenges due to variations in size, texture, and grayscale.
Purpose of the Study:
- To address blurred boundaries in liver tumor segmentation on CT scans.
- To overcome challenges in liver tumor classification caused by diverse tumor characteristics.
Main Methods:
- Utilized FLAS-UNet++ (UNet++ with fusion loss and atrous spatial pyramid pooling) for precise liver tumor segmentation.
- Implemented adaptive cropping and an improved Dense Block for enhanced feature extraction and classification.
- Integrated segmented tumor volume, patient demographics, and extracted features for a comprehensive diagnostic approach.
Main Results:
- Achieved Dice score of 71.9% and HD95 of 12.1 mm for liver tumor segmentation.
- Obtained classification accuracy of 82.4%, specificity of 79.8%, sensitivity of 84.4%, and AUC of 87.5%.
- Demonstrated superior performance compared to existing methods and mainstream networks.
Conclusions:
- The developed method enables accurate segmentation and classification of liver tumors in CT images.
- This approach holds significant potential for clinical application in liver tumor diagnosis and management.
