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

INTRODUCTION: Convolutional neural networks have begun to surpass classical statistical- and atlas based machine learning techniques in medical image segmentation in recent years, proving to be superior in performance and speed. However, a major challenge that the community faces are mismatch between variability within training and evaluation datasets and therefore a dependency on proper data pre-processing. Intensity normalization is a widely applied technique for reducing the variance of the data for which there are several methods available ranging from uniformity transformation to histogram equalization. The current study analyses the influence of intensity normalization on cerebellum segmentation performance of a convolutional neural network (CNN).

METHOD: The study included three population samples with a total number of 218 datasets, all including a T1w MRI data set acquired at 3T and a ground truth segmentation delineating the cerebellum. A 12 layer deep 3D fully convolutional neural network was trained using 150 datasets from one of the population samples. Four different intensity normalization methods were separately applied to pre-process the data, and the CNN was correspondingly trained four times with respect to the different normalization techniques. A quantitative analysis of the segmentation performance, assessed via the Sørensen-Dice similarity coefficient (DSC) of all four CNNs, was performed to investigate the intensity sensitivity of the CNNs. Additionally, the optimal network performance was determined by identifying the best parameter set for intensity normalization.

RESULTS: All four normalization methods led to excellent (mean DSC score=0.96) segmentation results when evaluated using known data; however, the segmentation performance differed depending on the applied intensity normalization method when testing with formerly unseen data, in which case the histogram equalization methods outperformed the unit distribution methods. A detailed, systematic analysis of intensity manipulations revealed, that the distribution of input intensities clearly affected the segmentation performance and that for each input dataset a linear intensity modification (shifting and scaling) existed leading to optimal segmentation results. This was further proven by an optimization analysis to find the optimal adjustment for an individual input evaluation sample within each normalization configuration.

DISCUSSION: The findings suggest that proper preparation of the evaluation data is more crucial than the exact choice of normalization method to prepare the training data. The histogram equalization methods tested in this study were found to perform this task best, although leaving room for further improvements, as shown by the optimization analysis.