Liver lesion changes analysis in longitudinal CECT scans by simultaneous deep learning voxel classification with...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Longitudinal contrast-enhanced CT (CECT) scans are crucial for monitoring liver lesions in oncology.
• Accurate identification and quantification of lesion changes are essential for clinical decision-making and treatment efficacy assessment.
• Current methods may be time-consuming and lack the precision needed for detailed longitudinal analysis.

Purpose of the Study:

• To develop and validate a fully automatic end-to-end pipeline for analyzing liver lesion changes in consecutive CECT scans.
• To introduce SimU-Net, a novel 3D R2U-Net model for simultaneous lesion identification and change detection.
• To enable robust longitudinal analysis of liver lesions, accounting for deformations and integrating previous segmentations.

Main Methods:

• A simultaneous multi-channel 3D R2U-Net model (SimU-Net) trained on registered CECT scan pairs.
• A model-based bipartite graph matching method for lesion-level change analysis.
• A longitudinal analysis method incorporating SimU-Net to handle liver deformations and previous segmentations.

Main Results:

• The SimU-Net pipeline achieved a mean lesion detection recall of 0.86 and Dice score of 0.82 for lesions > 5 mm.
• The system demonstrated superior performance compared to standalone 3D R2-UNet models, with approximately 50% higher precision.
• Lesion matching achieved 0.86 precision and 0.90 recall, while lesion classification showed 0.97 specificity and 0.86 recall.

Conclusions:

• The developed automatic pipeline provides accurate and comprehensive analysis of liver lesion changes.
• This approach has the potential to significantly reduce radiologists' workload and enhance radiological oncology evaluations.
• The SimU-Net model and associated methods offer a robust solution for longitudinal liver lesion monitoring in cancer patients.