Boundary-Sensitive Segmentation of Small Liver Lesions

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Hepatology

Background:

• Early diagnosis of liver diseases is critical for global health management.
• Small, early-stage liver lesions pose detection challenges due to indistinct boundaries and limited features.
• Existing methods struggle with the accurate segmentation and identification of tiny liver lesions.

Purpose of the Study:

• To develop an efficient model and a high-quality dataset for improved early liver lesion detection.
• To enhance the accuracy and fidelity of identifying small and inconspicuous liver lesions.
• To overcome limitations in feature extraction and boundary definition for small liver abnormalities.

Main Methods:

• Integration of path signature analysis with camouflaged object detection techniques.
• Development of a novel deep learning model leveraging path signatures to clarify lesion boundaries.
• Creation of a comprehensive dataset with over 10,000 liver images, including over 4,000 lesions, with a focus on small lesions.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art semantic segmentation and camouflaged object detection models.
• Significant improvements were observed in the detection of small liver lesions.
• Generated salience maps confirmed the model's robustness and accuracy in boundary region analysis.

Conclusions:

• The integrated model effectively addresses the challenges of detecting small liver lesions in medical imaging.
• The developed dataset and model provide a valuable resource for advancing liver disease diagnosis.
• The approach shows strong potential for clinical application in early liver disease detection and management.