Improving automatic liver tumor segmentation in late-phase MRI using multi-model training and 3D convolutional neural...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Hepatocellular carcinoma diagnosis benefits from dynamic contrast-enhanced MRI (DCE-MRI) due to higher sensitivity than CT.
• Most automatic liver lesion segmentation studies focus on CT, not DCE-MRI.
• Accurate liver tumor segmentation aids in planning liver interventions.

Purpose of the Study:

• To develop and evaluate a deep learning approach for liver tumor segmentation in DCE-MRI.
• To investigate the effectiveness of an anisotropic 3D U-Net and multi-model training for this task.

Main Methods:

• Anisotropic 3D U-Net architecture was employed for liver tumor segmentation.
• A multi-model training strategy was implemented to enhance segmentation performance.
• Segmentation accuracy was compared to a 2D U-Net and inter-rater agreement.

Main Results:

• The 3D U-Net achieved a mean Dice score of 0.70, outperforming a 2D U-Net (0.65).
• Multi-model training further improved the mean Dice score to 0.74, nearing inter-rater agreement (0.78).
• Expert evaluation showed 66% of segmentations were good or very good with multi-model training, versus 43% with single training.

Conclusions:

• Deep learning, particularly with multi-model training on DCE-MRI, enables accurate automatic liver tumor segmentation for detected lesions.
• Lesion detection, especially for smaller tumors, remains a challenge and requires further improvement.
• The developed method shows significant potential for clinical application in liver interventions planning.