Analysis of intensity normalization for optimal segmentation performance of a fully convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Convolutional neural networks (CNNs) outperform traditional methods in medical image segmentation.
• Data variability and preprocessing are critical challenges in CNN applications.
• Intensity normalization is a common technique to reduce data variance.

Purpose of the Study:

• To analyze the influence of different intensity normalization methods on CNN-based cerebellum segmentation.
• To investigate the sensitivity of CNNs to intensity variations in medical imaging data.

Main Methods:

• A 3D fully convolutional neural network was trained on 150 T1w MRI datasets.
• Four distinct intensity normalization techniques were applied to the data.
• Segmentation performance was quantitatively assessed using the Sørensen-Dice similarity coefficient (DSC).

Main Results:

• All normalization methods yielded excellent results on known data (mean DSC=0.96).
• Performance varied significantly on unseen data, with histogram equalization outperforming unit distribution methods.
• Input intensity distribution directly impacted segmentation performance, with linear adjustments optimizing results.

Conclusions:

• Proper preparation of evaluation data is more critical than the choice of training data normalization.
• Histogram equalization methods showed superior performance in this study.
• Individual data optimization offers potential for further improvements in CNN segmentation.