Magnetic Resonance Imaging Liver Segmentation Protocol Enables More Consistent and Robust Annotations, Paving the Way...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• High-quality datasets are crucial for AI in medical imaging, but dataset quality is often limited by inconsistent annotation.
• Lack of standardized labeling instructions hinders the development of reliable computer-assisted analysis tools for imaging examinations.

Purpose of the Study:

• To develop a standardized liver magnetic resonance imaging (MRI) segmentation protocol.
• To evaluate the impact of this protocol on annotation quality and inter-reader agreement in liver segmentation tasks.

Main Methods:

• Retrospective analysis of liver MRI scans from 20 patients with chronic liver disease.
• Manual segmentation by a radiologist and technician, followed by protocol development based on initial discrepancies.
• Second round of segmentation using the protocol, with Dice Similarity Coefficient (DSC) used to assess inter-reader agreement.

Main Results:

• Significant improvement in per-volume Dice Similarity Coefficient (DSC) for both T2-weighted imaging (T2wi) and T1-weighted imaging (T1wi) post-protocol implementation (p < 0.001 for T2wi, p = 0.03 for T1wi).
• Substantial enhancement in per-slice DSC for both imaging sequences (p < 0.001).
• Reduction in segmentations with non-annotated slices on T1wi (p = 0.04).

Conclusions:

• A standardized liver MRI segmentation protocol enhances annotation robustness and reproducibility.
• This improved data quality is essential for advancing AI applications in computer-assisted medical image analysis.
• The protocol's principles can be extended to other organs and lesions, potentially integrating into clinical guidelines for broader AI adoption.