Evaluating the impact of MR image harmonization on thalamus deep network segmentation

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Machine Learning

Background:

• Medical image segmentation is crucial for analyzing brain MRIs, aiding in tracking anatomical changes due to aging or disease.
• Contrast variations across datasets hinder consistent segmentation results from machine learning algorithms.
• MR image harmonization, an image-to-image translation technique, can standardize image intensity while preserving anatomical details.

Purpose of the Study:

• To present a 3D U-Net algorithm for segmenting the thalamus from multi-modal MR images.
• To investigate the impact of MR image harmonization on thalamic segmentation performance.
• To evaluate harmonization's effectiveness on a large dataset lacking ground truth labels.

Main Methods:

• A 3D U-Net algorithm was developed for thalamus segmentation.
• Two networks were trained: one on unharmonized MRIs and another on harmonized MRIs, using a dataset with manual labels.
• Performance was compared on a separate dataset with manual labels and on a large target dataset without labels.

Main Results:

• Networks trained on harmonized and unharmonized data showed no significant performance difference when evaluated on a similar dataset.
• However, the network trained on harmonized data demonstrated significant improvement when evaluated on the harmonization target dataset.
• Image harmonization effectively maintained anatomical integrity during intensity transformation.

Conclusions:

• MR image harmonization is vital for improving the accuracy of automatic segmentation of brain structures, particularly the thalamus.
• Harmonization enables the processing of large, multi-site MRI datasets, even without site-specific training data.
• This approach holds potential for broader applications in medical image analysis and understanding neurological conditions.