Fully Automated and Adaptive Intensity Normalization Using Statistical Features for Brain MR Images

Abstract

Accuracy of the results obtained by automated processing of brain magnetic resonance images has vital importance for diagnosis and evaluation of a progressive disease during treatment. However, automated processing methods such as segmentation, registration and comparison of these images are challenging issues. Because intensity values do not only depend on the underlying tissue type. They can change due to scanner-related artifacts and noise, which usually occurs in magnetic resonance images. In addition to intensity variations, low contrast and partial volume effects increases the difficulty in automated methods with these images. Intensity normalization has a significant role to increase performance of automated image processing methods. Because it is applied as a preprocessing step and efficiency of the other steps in these methods is based on the results obtained from the preprocessing step. The goal of intensity normalization is to make uniform the mean and variance values in images. Different methods have been applied for this purpose in the literature and each method has been tested with different kind of images. In this work; 1) The state-of-art normalization methods applied for magnetic resonance images have been reviewed. 2) A fully automated and adaptive approach has been proposed for intensity normalization in brain magnetic resonance images. 3) Comparative performance evaluations of the results obtained by four different normalization approaches using the same images have been presented. Comparisons of all methods implemented in this work indicate a better performance of the proposed approach for brain magnetic resonance images.

Research on melanoma image segmentation by incorporating medical prior knowledge

“…Comparing to the recent methods, the star priority-based melanoma segmentation method has further improved the performance of the segmentation model on the data set with fewer samples. An advantage of the proposed approach is that a separate intensity normalization stage ( Goceri, 2018 ), which usually leads to increase the computational complexity, is not needed.…”

Brain tumour segmentation framework with deep nuanced reasoning and Swin‐T

“…In the proposed approach, although the input images are noisy and include in‐homogeneous intensity values, there is no extra denoising method and intensity normalization such as in refs. [34–36], which may also cause increase in computational costs. $$$$\begin{equation} frelu(x)=relu(x+a)+b \end{equation}$$$$…”

Early Skin Disease Identification Using eep Neural Network

“…Capsule neural network 85.69% [54] Deep-learning Model 70-80% [55][56][57] 2021, 2018, 2019 Deep Learning 85% [58] Deep Learning 89% [59] CNN 80.2% [60] 5-layer CNN 95% [61] 14-layer CNN 97.78% [62] Distributed Learning System (DLS) 99.77% [63] Deep Residual Network 88.7%. [64] Local binary pattern (LBP) Resnet-50 and DenseNet-121…”