Medical Image Understanding

Medical Images

A medical image is an image that reflects the internal structure or internal functions of an anatomical area. It consists of a set of image elements-pixels (2D) or three-dimensional pixels (3D. Medical Images are characterized by discrete images produced by sampling or reconstruction, which map values to different spatial locations. The number of pixels is used to describe medical imaging under a certain imaging device. It is also an expression to describe anatomy and its functional details. The specific values expressed by pixels are determined by imaging devices, imaging protocols, image reconstruction, and post-processing.

Medical images have four key components: pixel depth, photometric representation, metadata, and pixel data. These components are related to the image size and resolution.
Image depth(Bit depth or color depth) is the number of BITs used to encode each pixel information. For example, an 8-bit grating can have 256 image depths ranging from 0 to 255.
Photometric RepresentationExplains how pixel data is displayed in the correct image format (monochrome or color image. To show whether color information exists in pixel values, we will introduce the concept of "number of samples per pixel. A monochrome image only has one "sample per pixel" and there is no color information in the image. Images are displayed in gray scale from black to white. The number of gray scale depends on the number of BITs used to store samples. Here, the gray scale is consistent with the pixel depth. Medical Radiation images, such as CT images and magnetic resonance (MR) images, are a gray-scale "photometric representation ". Nuclear Medicine images, such as positron emission tomography (PET) and single-photon emission tomography (SPECT), are usually displayed in color ing or color palette.
MetadataIs used to describe image information. It may seem strange, but in any file format, in addition to pixel data, the image also has some other information. Such image information is called "metadata". It is usually stored at the beginning of a file in the format of "data Header, covers image matrix dimensions, spatial resolutions, pixel depth, and light level representation.
Pixel dataIs the location where pixel values are stored. Based on different data types, pixel data uses numeric values to display the minimum number of bytes required, which are stored in the full or floating point format.

Image size = data header size (including metadata) + number of rowsPixel depth(Number of image frames)

Medical Image Format
There are 6 main formats of radiology images, including DICOM (medical digital imaging and communication), nifti (neural imaging information technology), and PAR/Rec (Philips magnetic resonance scanning), analyze (Mayo medical imaging), nrrd (near-original raster data), and mnic

Modern neuroimaging technology
EEG, single-photon emission layer imaging (SPECT), positron emission tomography (PET), functional magnetic resonance imaging (fMRI), invasive optical imaging (invasive optical imaging ), intracranial recording and ECoG ). Among them, the most widely used are fMRI and pet.

Mni SpaceIt is a coordinate system established by the Montreal Neurological Institute based on a series of normal human brain magnetic resonance images.
Native space is the original space, and the image has not been processed. At this time, the images of different subjects are not comparable. Therefore, all the images of the subjects must be registered and standardized to the same template, in this way, all tested dimensions, origins, and voxel sizes are the same. Using the mni standard template, the image is converted to The mni space.
Brain Imaging data mainly consists of three modes: Mhd, fMRI, and 3D. Among them, the user can use one of the three dimensions (1 dt1) and the user can use one of the four dimensions (4-dimension ).

• Diffusion: diffusion tensor imaging, magnetic resonance diffusion tensor imaging
  
• FMRI: functional magnetic resonance imaging, functional magnetic resonance imaging
  
• DICOM (Digital Imaging and communications in medicine) is an international standard for medical digital imaging and Communication (ISO 12052 ). It defines the medical image formats that can be used for data exchange that meet the clinical needs of Quality

PET is short for Positron Emission Tomography (positron emission tomography), an advanced nuclear medicine imaging technology; CT is short for computed tomography, x-ray tomography is a widely used and rapidly developing X-ray tomography technique. These two technologies are organically integrated into the same device, and different types of images are displayed on the same machine, forming a PET/CT

Various types of image details MRI

MRI imaging is characterized by a variety of imaging sequences. These imaging sequences can produce distinctive MRI images that not only reflect human anatomy, but also reflect physiological function information such as human blood flow and cell metabolism.
MRI scan methods can be divided into regular scan and functional scan. Regular scans mainly reflect anatomy, while functional scans reflect the body's metabolism, blood flow, and other functional information in different ways. Conventional scans include T1 weighted, T2 weighted imaging, angiogram imaging, and dynamic enhancement imaging. Functional Imaging includes diffusion-weighted imaging (SNR), perfusion-weighted imaging (PWI), magnetic resonance pop imaging (MRS), and blood oxygen saturation level-Dependent imaging (BOLD ).

T1 weighted Highlight shows the anatomical structure, T2 weighted can highlight the lesions

MRI is the most common and important method for checking brain lesions. Compared with CT, MRI has no bone artifacts and has better soft tissue resolution. Additionally, you can flexibly select the axis, crown, vector, and oblique position scans as needed to fully display lesions.
MRI is applied to multiple aspects of cardiac examination, including morphological examination, high-resolution imaging of cardiac anatomy, cardiac functional Examination, evaluation of shot blood scores EF, per stroke output SV, late-contraction and late-relaxation volume, cardiac output, and valve properties; myocardial perfusion and myocardial activity; coronary anatomy and blood flow; Myocardial Metabolism; high-resolution arterial wall plaque Imaging
Generally, t1 and t2 scan sequences are used in spine examinations, and multiple scan sequences are generated at the vector, axis, coronary, and any angle.
MRI is one of the commonly used methods for the diagnosis of prostate hyperplasia and prostate cancer. Generally, t1 and t2 scans are used, and the most commonly used Scanning direction is horizontal axis. In addition, dynamic contrast-enhanced scans, diffusion-weighted imaging (SNR), and magnetic resonance spectroscopy (MRS) are also used)
For the liver, t1 and t2 scans and dynamic enhancement scans are usually used. Dynamic Enhanced scanning technology should be considered as a conventional method for liver, especially Liver Cancer Examination

FMRI

FMRI Principle
Oxygen is required for nerve cell activity, which is carried by hemoglobin in The microvessels around the nerve cells. Therefore, when the nerves are active, the nearby blood flow will increase to supplement the oxygen consumed in time, and eventually increase the local blood, changes the concentration of oxygen-containing hemoglobin and oxygen-oxidized hemoglobin in the blood. Changing the concentration of oxyhemoglobin is a kind of magnetic resonance material, which may lead to changes in the signal strength of magnetic resonance. The fMRI scanner samples this continuously changing magnetic signal at a certain time resolution, and finally obtains a time sequence that can reflect the activity of neural cells (Meta. In nuclear magnetic resonance medicine, changes in magnetic resonance signals caused by changes in oxygen hemoglobin and oxygen hemoglobin levels are called the blood oxygen level dependence (BOLD) effect, and the corresponding magnetic resonance signal is also called the BOLD signal
MRI data features

• Low Signal-to-Noise Ratio
  
• High data dimension
  
• Large data distribution difference

MRI and MRI
MRI scans the structure image of the brain, also known as the T1 weight image. It has a very high spatial resolution, from which you can see a very clear anatomical structure, you can also distinguish between a variety of different organizations
FMRI is often used to study the specific functions of the brain. Functional Images are scanned, also known as T2 * weight images. Although its spatial resolution is relatively low, but the time resolution is very high, you can scan a stack of functional images in a short time. In this way, we can study how experimental operations affect the brain's MRI signal.

FMRI data preprocessing
Data preprocessing steps include: Visualization, artifact removal, slice time correction, and motion correction) correction for physiological effect, co-regiation, normalization, and spatial and temporal filtering)

Main steps of fMRI Analysis

1. Quality control: ensures that data is not damaged by false traces
   
2. Distortion Correction: Correction of space distortion that often occurs in fMRI images
   
3. Head movement correction: Correction head movement, re-aligning the scanned Time Series Image
   
4. Inter-layer time correction: time difference between different layers of the image is corrected.
   
5. Spatial standardization: data of different individuals is aligned to a general spatial structure, so that all data can be combined for group analysis.
   
6. Space smoothing: intentionally blurred data to reduce noise
   
7. Time Filtering: filters data in the time dimension to remove low-frequency noise.
   
8. Statistical modeling: fitting the statistical model to the observed data to estimate the response caused by a task or stimulus
   
9. Statistical Inference: the statistical significance of the estimation results, and correction for a large number of statistical tests carried out throughout the brain
   
10. Visualization: visualize the results and estimate the effect volume

Simplified procedure:

1. Alignment: the error caused by human factors is very serious in the MRI sequence, such as small movement of the head. Therefore, we should minimize these human factors before processing data.
   
2. Standardization: due to differences between individuals, when extracting the mean signal between individuals or describing the activation zone under the coordinate system of the standard space, the images of many individuals need to be deformed to the same standard space, that is, image space standardization. Currently, standard brain maps of Talairach and tournoux are most commonly used.
   
3. Registration: to precisely locate the function activation area, the function information is usually located on an anatomical image with a higher resolution (such as a T1 weighted image). This requires registration of functional and anatomical images.
   
4. Smooth: There are two main functions of smoothing: increasing the signal-to-noise ratio and enhancing partial image effects.

Three major coordinate axes Based on MRI standard coordinate space
In the standard space used for neural imaging data, X represents left/right, Y represents front/back, and Z represents top/bottom. In the data matrix, a specific element can be marked as [xvox, yvox, zvox]. The coordinates of these three dimensions can be used to determine the position of the element. As follows:

DICOM

It defines medical image formats that meet clinical needs for data exchange and can be used to process, store, print, and transmit medical image information. DICOM can be easily exchanged between two workstations that meet the DICOM protocol.
A dicom file consists of a Data header and image data. The size of the Data header depends on the amount of data information. The data header contains the patient id and patient name. It also determines the number of frames and resolution.

One scan of each patient may contain dozens or more than one hundred DCM data files (slices ). It is suffixed with. DCM and can be read using Python DICOM package. Its pixl_array data is generally used.

DICOM format data processing process
Scan is actually a three-dimensional image. After reading the code, you can open-source slices to view slices of different slices.
Secondly, the CT scan map contains all the organizations. If you look at it directly, you cannot see any useful information. Some preprocessing is required. An important concept in preprocessing is radiation dose measured in Hu (hounsfield Unit). The following table lists the tissue and organ corresponding to different radiation doses.

Substance	Hu
Air	-1000
Lung	-500
Fat	-100 ~ -50
Water	0
CSF	15
Kidney	30
Blood	30 ~ 45
Muscle	10 ~ 40
Gray Matter	37 ~ 45
White Matter	20 ~ 30
Liver	40 ~ 60
Soft tissue, contrast	100 ~ 300
Bone	700 (cartilage )~ 3000 (Cortex Bone)

Calculation method:

Hounsfield Unit = pixel_value * rescale_slope + rescale_intercept

Generally, rescale slope = 1, intercept =-1024
The gray value is the value that pixel value is converted to for display after many LUTs, and this conversion process is irreversible. That is to say, the gray value cannot be converted to the CT value. You can only obtain an approximate range based on the window width.

A simple DICOM Image Processing

Import dicomimport pylabds = DICOM. read_file ("1.2.840.113619.2.55.3.2831194257.596.1285460208.412.1.dcm") print ("image attributes:", DS. dir ("Pat") print ("Patient:", DS. patientname) # CT worthy matrix pix = Ds. pixel_array # Read the image pylab. imshow (Ds. pixel_array, cmap = pylab. cm. bone) pylab. show () # set the value smaller than 300 in the image to 0for N, Val in enumerate (Ds. pixel_array.flat): If Val <300: Ds. pixel_array.flat [N] = 0ds. pixeldata = Ds. pixel_array.tostring () ds. save_as ("output. DCM ")

Source image:

Display chart:

Figure after filtered value:

Nifti format

One of the major features of NiTi is that it contains two definitions of affine coordinates, which can be used to define the index (I, j, k) and spatial position (X, y, Z ).
The extension of the standard nifti image is (. NII), which contains header files and image data. At the same time, nifti can also use independent image files (. IMG) and header files (. HDR)
Differences between DICOM and nifti
The main difference between DICOM and nifti is that original image data in nifti is stored in the format of 3D images, while DICOM is stored in the format of 3D image fragments. That's why in some machine learning applications, nifti is more popular than DICOM because it is a 3D image model. Processing a single nifti file is much easier than processing hundreds of DICOM files. Each 3D image of nifti only needs to store two files, but more files in DICOM.

Pet

The unique role of PET is based on Metabolic Imaging and quantitative analysis. It is used as a tracer to use short-lived NNS, such as 11C, 13N, 15o, and 18f, which constitute the main elements of the human body, it not only can quickly obtain multi-level tomography, three-dimensional quantitative results, and three-dimensional full-body scans, but also can dynamically observe the physiological and biochemical changes of metabolites or drugs in the human body from the molecular level, it is used to study human physiological, biochemical, chemical transmitter, receptor, and gene changes. In recent years, pet has shown its unique superiority in diagnosing and guiding the treatment of tumors, coronary heart disease and brain diseases.
PET is the abbreviation of positron emission tomography. The clinical imaging process is to mark positron-emission radiomns (such as F-18) to compounds that can be involved in human tissue blood flow or metabolic processes, inject the radiospector labeled with positron compounds into the subject. Allow the subject to perform PET imaging within the effective field of view of PET
The basic principle of CT is image reconstruction, which varies with X-ray absorption from various human tissues (including normal and abnormal tissues, after dividing a selected layer of the human body into many small cubes (also called body elements) and X-rays pass through the body elements, the measured density or gray value is called a pixel. The X-ray bundle passes through the selected layer, and the detector receives the sum of the attenuation values after the X-ray absorption by each body in the direction of the X-ray bundle, which is a known value, the X-ray attenuation value of each element that generates the total amount is unknown. When the X-ray source and detector are moving around the human body in an arc or circular circle. The X-ray attenuation value of each element is obtained by iterative method and the image is reconstructed to obtain the black and white images with different density at the layer.
PET-CT perfectly integrates pet and CT, and pet provides detailed functional and metabolic molecular information of lesions, while CT provides precise anatomical location of lesions, one-time imaging can obtain the fault images of all directions of the whole body, which is sensitive, accurate, specific, and precise to clearly understand the overall situation of the whole body, so as to achieve early detection of lesions and disease diagnosis.

References:
Python and Medical Image 2
Common medical scan image processing steps
Medical Image Analysis in deep learning 2
MRI image learning notes
Medical datasets and machine learning projects
Functional division of the human brain for fMRI data

Medical Image Understanding