Multi-Center MRI Carotid Plaque Component Segmentation Using Feature Normalization and Transfer Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Automated segmentation of carotid artery plaque components in MRI is crucial for studying plaque vulnerability and clinical applications.
• Supervised classification methods show promise but struggle with data from different scanners, necessitating reduced manual annotation.
• Widespread implementation requires methods robust to variations in imaging data.

Purpose of the Study:

• To segment carotid plaque components (fibrous tissue, lipid tissue, calcification, intraplaque hemorrhage) in multi-center MRI data.
• To compare traditional same-center training with approaches using limited or no same-center annotated data.
• To evaluate the efficacy of feature normalization and transfer learning for cross-dataset performance.

Main Methods:

• Voxelwise tissue classification was performed using traditional same-center training as a reference.
• Two novel approaches were evaluated: nonlinear feature normalization and transfer learning algorithms.
• These methods utilized both same-center and different-center annotated data with varying weights.

Main Results:

• The combination of feature normalization and transfer learning yielded the best segmentation results.
• Unlike other methods, the proposed approach showed no significant differences in voxelwise or mean volume errors compared to same-center training.
• Significant differences were observed for other approaches when compared to the reference method.

Conclusions:

• Extensive feature normalization and transfer learning are valuable for developing robust supervised segmentation methods.
• These techniques enable supervised methods to perform well across diverse datasets, including multi-center studies.
• The proposed combined approach addresses the challenge of data variability in carotid plaque MRI analysis.