Liver segmentation and metastases detection in MR images using convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Hepatobiliary Diseases

Background:

• Liver metastases are common complications of primary tumors, significantly impacting patient prognosis.
• Early and accurate detection of liver metastases is critical for effective treatment planning and improved patient outcomes.
• Current detection methods can be limited, necessitating advanced diagnostic tools.

Purpose of the Study:

• To develop and validate a deep learning-based method for automated liver segmentation and subsequent detection of liver metastases.
• To evaluate the performance of convolutional neural networks (CNNs) using multi-modal MRI data for liver metastasis detection.
• To assess the clinical utility of combining dynamic contrast-enhanced (DCE) and diffusion-weighted (DW) MRI sequences for enhanced diagnostic accuracy.

Main Methods:

• Automatic liver segmentation was performed using six phases of abdominal dynamic contrast-enhanced (DCE) MRI.
• A dual-pathway convolutional neural network was employed for metastases detection within the segmented liver mask.
• The network integrated information from both DCE-MR and diffusion-weighted MR images.

Main Results:

• The automated liver segmentation achieved a high median Dice similarity coefficient of 0.95 compared to manual annotations.
• The metastases detection model demonstrated a high sensitivity of 99.8%.
• The detection method yielded a median of two false positives per image, indicating robust performance.

Conclusions:

• High-quality liver segmentation is achievable with the proposed automated method.
• The developed CNN-based approach effectively detects liver metastases with high sensitivity.
• The integration of DCE-MR and DW-MR sequences in a dual-pathway network significantly enhances the detection of liver metastases.