paper: https://link.springer.com/article/10.1007/s10278-017-9983-4

Abstract

Accurate brain MRI segmentation of structures is relied on for neurological diseases and ailment. The interest on deep-learning segmentation method has gained interest all around because of their learning and skill to process large amount of data, and with their architecture it gradually surpasses the prior state of the art machine learning model. This paper talked about the deep-learning approach for quantitative brain MRI and how it is used for brain lesion segmentation. Also, the performance, the speed and the properties were discussed. The patch-wise CNN, Semantic-wise CNN and the cascaded CNN was proposed but the cascaded convolutional neural network had a better performance than other methods which approach on an unseen test set was 93.3% (sensitivity) and 82.22% (specificity). Finally, a valuation of the current state of the art and also recognize future works.

Introduction

MR imaging is the most common choice for structural brain study, since soft tissues have high contrast and high spatial resolution, its risk free. Brain analysis of MR image is widely used for brain illnesses description such as Alzheimer, epilepsy disease, schizophrenia disease, cancer disease, and infectious and deteriorating. Segmentation, which entails labeling 2D pixels or 3D voxel is an important piece of brain analysis. Although manual segmenting is a standard in-vivo image but however this involves outlining structures of slice-by-slice, and it’s tedious and not error free. These brings about the automation of segmentation technique and it provides high accuracy. With 3D and 4D images becoming more popular in the medical environment, medical images are increased in size and complexity. Therefore. Developing tool to assist with extracting information from these datasets are essential and machine learning algorithm allows computers to make data-driven predictions from large datasets.

Methods

Datasets

Brain Tumor Segmentation (BRATS)

Ischemic Stroke Lesion Segmentation (ISLES)

Mild Traumatic Brain Injury Outcome Prediction (mTOP)

Multiple Sclerosis Segmentation (MSSEG)

Neonatal Brain Segmentation (NeoBrainS12)

MR Brain Image Segmentation (MRBrainS)

Brats: This brain challenge is in conjunction with MICCAI (Medical Image Computing and Computer Assisted Intervention) which holds annually and dates back to 2012 which compares the start of the art in automated segmentation  of brain tumor. A large dataset is used with five labels which are healthy brain tissue, necrosis, edema, non-edema, non-enhanced, then also enhanced regions of a tumor)

Isles: This challenge is to evaluate stroke lesion / clinical outcome prediction from acute magnetic resonance images scans. A large number of acute stroke cases then associate clinical parameters are provided.

mTOP: This challenge is about finding the difference between a healthy brain and a traumatic brain patient and the data is sort in an unsupervised manner.

MSSEG: This challenge is focused on the state of the art and advanced methods of segmentation for the participating data. This challenge evaluates but the lesion detection and lesion segmentation on a multiple data base of 38 patients from different centers.

NeoBrainS12: This challenge is to associate segmentation algorithms of a neonatal brain tissues and measurement of matching volume with the use of T1 and the T2 brain magnetic image scan. This compares the cortical and central gray matter, non-myelinated and myelinated white matter, brainstem and cerebellum and cerebrospinal fluid in the ventricles and in the extracerebral space.

MRBrainS: This challenge aims to assess the segmentation algorithm for gray matter, white matter and cerebrospinal fluid on multi-sequence T1 weighted, T1-weighted-inversion recovery, and FLAIR), these includes 3 Tesla scans of the brain.

Training, Validation and Evaluation: data are divided into three sets which are (training, validation and testing) in machine learning. Furthermore, machine learning algorithm learns from examples, provides good learning results and also the ability of an established algorithms of hidden data. Some other methods are preferred with the likes of one-leave out, fivefold or tenfold validations, when there is limited data.

Image preprocessing: Before automated analysis, there are some steps essential for images to look more familiar and these are steps are stated to as pre-processing. Also, the MRI images are difficult due to its inhomogeneity, variability of the intensity range and noise.

Registration: Image registration is the procedure of converting different sets of data into one synchronize system. The images are taken from several sensors at different times and at multiple view-points.

Skull Stripping: These is one of the initial steps in detecting abnormalities in the brain. It is the process of isolating brain tissue from non-brain tissue in an MRI image of a brain.

Bias Field Correction: This is an improvement of the image contrast differences because of its magnetic field inhomogeneity.

Intensity Normalization: This is an important pre-processing stage in magnetic resonance image (MRI) analysis. Throughout magnetic resonance image acquisition, different scanners would be used for scanning different subjects or the same subject at a different time, which large intensity variations can occur has a result.

Noise Reduction: Decrease of the locally-variant Rician noise detected in magnetic resonance images. Rician noise makes image-based quantitative dimension tough.

CNN Architecture Methods.

Patch-wise CNN Architecture: This approach trains convolutional neural network for segmentation. A patch is extracted from a given image around each pixel and this model is trained and given a class to identify a normal brain or a tumor brain. This design has multiple convolutional, activation functions, max pooling and the full convolutionally layers serially. Hence, the outcome of the neural network and the trained model is correctly identified by the given class label

Illustration:

Semantic-Wise CNN Architecture: This design makes forecasts the entire input image of each pixel of like semantic segmentation. They are similar to autoencoder in which they encode parts that extracted features and decodes the part that up samples the higher-level features from the encoder part and combines lower level features from the encoder part to classify pixels.

Illustration:

Cascaded CNN Architecture: This model joins two CNN design, where the output of the first convolutional neural network is regarded has the input which is used to get classification result and also the first convolutional neural network is used has the training set while the with second convolutional neural network is used to tune the first convolutional neural network.

Illustration:

Segmentation of normal brain structure: Correct automatic brain magnetic resonance image segmentation is important in the study of infants and quantitative valuation of the brain tissue and intracranial volume in large scale studies.

Segmentation of Brain Lesions: Brian examination of lesion include measuring the imaging biomarker such as volume, progression of quantify treatment of said disease, such like stroke and cancer of the brain. Dependable feature extraction of these biomarkers depends on previous correct segmentation. Several automated techniques have been proposed for segmentation of lesion, which includes unsupervised learning model which aims to automatically adapt new image data, also supervised learning technique which for a give dataset, it learns the textural and appearance properties of the lesion, also the supervised learning and unsupervised learning is combined by deep learning into an integrated pipeline by registration of labeled data into a common anatomical space.

RESULTS

Deep Learning approaches for brain structure segmentation

Deep Learning approaches for quantification of brain lesions

Related work

Alexandre de Brebisson  “Deep Neural Networks for Anatomical Brain Segmentation”

This paper takes a step of segmenting human brain magnetic resonance (MR) images automatically into anatomical regions.  Deep neural network is the method that gives each voxel in an MR image of the brain to its matching anatomical region. Information are captured on different scales from the input image at different scales around the voxel of interest.

Conclusion

Having applied deep learning methods to MRI brain segmentation, it shows they outperform prior state of the art machine learning method. Due to compound anatomy of the brain and brain variability presence, non-standardized magnetic resonance scales due to variability in imaging procedures, image gathering deficiency, and existence of pathology, analysis has been a great task to computer-aided methods. Deep learning is a more general method that can handle these variabilities. To learn, it is important for Image to go through preprocessing. Various pre-processing stages have been applied to improve learning process. The aim of a learning technique is having a flawless classification, nonetheless there are areas in an image that overlap amongst classes, this is called partial volume effect.

Future Work

To improve of deep learning architectures for the future, Adaptation of a good deep learning network that is trained on a lager dataset.