Deep Learning Framework for Liver Segmentation from T₁-Weighted MRI Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Liver segmentation in radiological images is challenging due to anatomical variability.
• Magnetic Resonance Imaging (MRI) offers superior soft tissue contrast for liver pathology diagnosis compared to CT scans.
• Automatic segmentation of liver from MRI is difficult due to the absence of Hounsfield unit-based preprocessing.

Purpose of the Study:

• To investigate state-of-the-art segmentation networks for liver segmentation in volumetric MRI.
• To develop and evaluate a novel cascaded network for accurate liver segmentation from MRI slices.

Main Methods:

• Utilized T1-weighted (in-phase) MRI scans from the CHAOS dataset (20 patients, 647 slices).
• Evaluated twenty-four diverse state-of-the-art segmentation networks with various encoder/decoder backbones.
• Proposed and implemented a novel cascaded network architecture for axial liver slice segmentation.

Main Results:

• The proposed cascaded network achieved a Dice Similarity Coefficient (DSC) of 95.15%.
• The network obtained an Intersection over Union (IoU) score of 92.10%.
• The developed framework demonstrated superior performance compared to existing methods on the same test set.

Conclusions:

• The novel cascaded network provides highly accurate liver segmentation from MRI.
• This automated approach can significantly enhance computer-aided diagnosis for liver diseases.
• The method offers a robust solution for challenging MRI-based liver segmentation tasks.