Medical Image Analysis Flashcards
What are the assumptions of pixels?
They are square and non-overlapping
What are the 3 image types?
Intensity images Colour images Binary images
What is an intensity image?
Each pixel has a similar intensity This is the most common type of image Value is proportional to intensity e.g. value proportional to dose/light Often referred to as a greyscale image No need to stick to shades of grey
What is a colour image?
Each pixel has a colour value Need a colour model
What is a binary image?
Each pixel has only 2 possible states 1 or 0, on or off, true or false Used in medical image analysis Used in the processing stage Not helpful to the end user
What is the problem with 2D greyscale images?
The image is not equally sharp due to movement during exposure
In a 2D colour image what do the pixel value represent?
Luminance Brightness Colour Using a colour model
What is the most common colour model?
R-G-B
What is the RGB colour model?
Most common Additive colour model Start with black, add Red-Green-Blue The pixel value has 3 intensities for RBG light
What are the RGB components in a white pixel?
RGB are equal and maximum
Describe the dimensions in a 3D image
N x M x D Third dimension can be spatial, temporal or some other variable Spatial - CT scanner volume Temporal - fluoroscopy
What are binary images used for?
they are often used to identify objects or regions within an image Used for image processing Set a threshold intensity to t. If the intensity is greater than t then 1 or on and vice versa. Can have 2 threshold values, between x and y
What are the basics behind displaying intensity images?
There must be a map from the pixel value to the colour of the pixel to be displayed.
What is the typical computer display?
RGB colour device with 256 possible shades of each colour
Describe the RGB values in a grayscale image?
If the pixels are grey, black or white then the values for each channel are equal 0,0,0 is black, maximum in each channel is white
What is the possible colour range display in a 24 bit image?
256 red x 256 green x 256 blue= 16 million colour range display Each colour has 8 bits assigned to it 8 bits = 2^8 shades = 256
What is the problem with displaying medical images?
There are often more pixels than is available on a standard screen They often contain many more distinct intensity values and so mapping is required
What is a LUT?
Look up table It is used as a map between pixel value and the colour to display It may produce a colour or greyscale representation of an image If a LUT has fewer entries than in the image, the pixel values in the image must be scaled to LUT entries
What is a true colour image?
They are capable of being displayed directly. This means that each pixel value has a distinct colour. 16 million possible colour shades. If there are less colours available than are recorded in the image then there is scaling.
What are the advantages of using a colour LUT?
Allows you to identify areas of similar intensity Can draw attenuation to an area difficult to see in grey
What are the 2 methods of number representation?
Integer Floating point
How are images stored on a computer?
The images pixels are represented in the computers memory or storage devices Computer representation of numbers has its limitation Everything in a computer is represented in binary
What is binary>
It is a base number system and is either 0s or 1s Binary numbers: Units 2s 4s 8s 16s It is much more efficient to count in binary
What is a bit?
Binary digital Smallest amount of data a computer can represent It is either a 0 or a 1 b
What is a byte?
8 bits 8 0s or 1s It is the basic unit of storage on a computer B
What is a word?
It is the number of bytes native to the design of the computer hardware or OS Modern computers typically use a 4 byte word (32 bit) or 8 byte words (64 bit)
What is the consequence of using larger words in computer storage?
Using larger words increases memory
What is the binary convention for binary multiples
1024 = kibi = Ki 1024^2 = mebi = Mi 1024^3 = gibi = Gi 1024^4 = tebi = Ti 1024^5 = pebi = Pi 1024^6 = exbi = Ei
What is 1MiB in bytes?
1024^2 = 2^20 bytes
What is 1MB in bytes?
1 x 10^6 bytes
How are integers stored on a computer?
They are represented as fixed length block of bits Can be signed or unsigned
What is a signed integer?
It can hold negative and positive whole numbers It shifts the range to allow for the storage of negative values
What is an unsigned integer?
It can only hold numbers that are greater or equal to 0
How many possible states does a byte have?
1 byte = 8 bits = 2^8 possible states = 256
What is the range of unsigned byte (integer)?
0-255
What is the range of a signed byte (integer)?
-128 - 127 7 number bits and 1 sign bit
How does a computer store an signed integer?
Early computers used a dedicated sign bit (+ or -) Modern computers use a system called two’s complement to represent the negative values
What is two’s complement
It is used to store negative integer values For an N bit number, the two’s complement is calculated by subtracting the number from 2^N It allows for efficient add, subtract and multiply operations
What effect does increasing the storage bits have on the range of values a pixel can take?
Increase bits to the storage used for an integer gives a larger range
What are the common integer sizes?
8, 16, 32, 64 and 128 bits
Describe the integer lengths in a 32 bit computing environment
16 bit integers are short integers - half word size 32 are integers - word size 64 are long integers - double word size
What happens if a pixel value goes over the top of the range it can take?
Overflow occurs. Any pixels that appear over the maximum will appear as the maximal colour, or potentially wrap around to the minimum
What are the pros and cons of integer representation?
It is not good for mathematical equations or processing Unsigned is OK for X-ray as the dose cannot be negative It has a limited number of values it can take Can’t do mathematical equations with fractions It can only be a whole number and can’t represent fractions Quicker than floating point
What are floating point numbers?
They divide their storage into a sign bit (S), an exponent (E), and a fractional part (F) in order to represent a number V = -1^S x 2^E x (1.F) The number of bits in each section is dependent on the word length
What are the pros and cons of floating point numbers?
It allows numbers of differing magnitude to be recorded with reasonable precision Allows you to store non-whole numbers reasonably accurately. You do lose some detail - the importance of this is dependent on the application Slower than integer Very few numbers can be represented exactly as floating point
In floating point what is the consequence of increasing the number of bits available?
Increases the precision of the numbers that can be stored
Why would you use compact data types?
Decrease the storage size requirements and increase transmission
What are the standard formats of storing images?
JPEG (JPG) Tagged Image File Format (TIFF) Windows bit map (BMP) Portable Network Graphics (PNG)
What are the limitations of using JPEGs for medical images?
They are compressed to take less disk space Irreversibly degrades the quality of the image 8 bit
What are the advantages of using TIFF for medical images?
More flexible but there are problems with compatibility Can do 3D and 16 bit images
What is the disadvantage of using BMPs for medical images?
Limited to 8 bit images
What are the components of a medical image?
Need pixel intensities as well as other features: Who, when, where, scan settings
What does DICOM stand for?
Digital Imaging and Communications in Medicine
What is DICOM?
A standard for storing and communicating medical images, it overcomes some of the issues seen with other file formats. Can store a wide range of information relevant to the image which can cannot be stored in other formats: patient details, scan settings, ECG traces etc. It is a complex standard and there are issues with compatibility of files between devices/software.
What are the 2 resolution components in medical imaging?
Spatial resolution - refers to pixel pitch Greyscale resolution refers to how finely divided the greyscale is.
What does the resolution control?
It controls the maximum amount of information an image can hold and also determines the subtly of detail that it is possible to represent.
What happens if you decrease the pixel pitch?
Decrease the distance between the pixel centres Image will contain more pixels if the same area is imaged Finer detail will be visualised More detail in the image - increase resolution Decreased signal per pixel Decreased signal to noise ratio
What happens if you increase the pixel pitch?
Decreased spatial resolution Less detail Increased unsharpness
What happens if you decrease the pitch by 1/2?
4x the information and increase the storage requirements
What is needed to gauge spatial resolution?
Number of pixels Area covered by the pixel matrix
What factors can affect spatial resolution?
Pixel pitch Scanner specification (e.g. scintillator thickness, focal size) Scanner settings (SOD, SID, selected focus) Patient movement
When is interpolation used?
Used when re-sampling the pixel matrix of an image e.g. when increasing the size of an image for display purposes Interpolation improves the cosmetic appearance of the image but IT DOES NOT ALTER RESOLUTION
What are the methods of interpolation?
Nearest neighbour Bilinear interpolation Bicubic interpolation
What is nearest neighbour?
This substitutes the vale of the closes existing pixel centre of the new pixel centre location Images: checkered background with blocky edges
What are the advantages of nearest neighbour?
Computationally efficient Fastest method
What are the disadvantages of nearest neighbour?
Pixel blocking is common Poor edge representation Degraded image quality
What is bilinear interpolation?
This takes the weighted average of the neighbouring pixel values and this is used to calculate the new pixel value. The nearest 2x2 neighbourhood to the desired pixel is used to form the new value
What are the pros and cons of bilinear interpolation
More accurate More time consuming Better appearance to the image
What is bicubic interpolation?
The nearest 4x4 neighbour hood to the desired pixel location is used to calculate the new pixel value. A cubic polynomial surface is generated for the values to generate a smooth interpolation
What are the pros and cons of bicubic interpolation?
It is mathematically more complex but gives little advantage over bilinear.
What is the effect of interpolating an image?
It smooths the image and removes the blockiness The innate resolution of the image is not changed. Improvement is cosmetic.
How are analogue signals digitised?
Analogue to Digital Converter (ADC)
How does an ADC operate?
It will operate on a specified range of input values and produce digital values divided into a finite number of digital values Using a finer greyscale allows more subtle differences in contrast to be appreciated.
How is a finer greyscale achieved?
Sampling the same init range with more discreet levels This will improve greyscale resolution
What are histograms and what are they useful for?
They are useful when analysing the distribution of an images pixel intensities. Plots all possible intensities (x) against the number of pixels at each intensity (y). It can be useful to determine if the exposure is correct.
How will an overexposed and underexposed image appear on a histogram?
Over exposed - all of the values are to the right (highest) end of the histogram VV for under exposed Need to look at the distribution of values in the histogram
What does segmentation involve?
It involves identifying and outlining structures within an image The result is a binary image
What is the result of segmentation?
It is a 2D binary image which represents an anatomical structure. It can be called a mask
What operations can be applied to a segmented image?
OR (merge 2 images together) AND NOT (creates reverse image)
What is the purpose of defining a region of interest in an image?
Statistical analysis Used to calculate average intensities Can also be used to create a volume Visualisation
How do you calculate the volume of a ROI?
Number of voxels x volume of the voxel
Describe manual segmentation
It is done via boundary tracing or using predefined shapes or area painting
What are the advantages of manual segmentation
Simplest Most common
What are the limitations of manual segmentation?
Very slow - need to be able to justify the time and resultant expense Interobserver variability - different people get different results This is less important in a clear image - could be significant diagnostically
What are the methods used for automated/semi-automated segmentation?
Thresholding Grouping Region growing Boundary methods
What is thresholding?
The principle is to select all values that lie within a given intensity range This can be combined with a rough manual outline It makes the process less user dependent
How can thresholding be done?
By setting a range of accepted values Edge detection - depicts the steepness of signal changes. Apply a threshold to the gradient Histogram - Makes it more reproducible and it allows a more objective cut off between dark and light
What is grouping?
Thresholding with more dimensions It is thresholding on 2 images at the same time and is much more complex. It loses all anatomical reference You draw a ROI in a space where 2 images are represented together.
How is region growing done?
- Set a seed point manually 2. Add the neighbouring pixels with a similar intensity 3. For every new pixel, repeat 2 and so on until no more new pixels are found
What is region growing?
Automatic algorithm that checks if surrounding pixels are of a similar intensity and only adds them if they are. Need to define what similar is. Generally defined as +/- X% Minimal user dependence
When is region growing useful?
When there is a dark object on a light background or VV
What are the boundary methods used in segmentation?
Isolines and isosurfaces look for lines in the image with similar intensities Active contours - grow a boundary into a shape Watershed, Neural networks, model based, atlas guided, fractal methods, merging, artificial intelligence
What is visualisation?
Only a problem in digital imaging. Energy is converted into a pixel value and visualisation is the process of viewing these pixel images as a value Visualisation depends on the mechanism used
What are the methods for 2D visualisation?
Grey scales Colour coding Plots and profiles Fusion
What is the principle behind greyscale visualisation?
Paint by numbers Image is series of numbers Look at the range of pixel values and create a greyscale image Need to select a maximum and minimum value Greyscale is then fitted to this range Determined by a LUT
What needs to be selected in greyscale visualisation?
Maximum value Minimum value How the scale changes from white to black
What happens as you decrease the maximum value on a greyscale?
Increase brightness/contrast
What is the level>
Brightness Centre of the range
What is the window?
Contrast Width of the range
What is the consequence of changing a greyscale to predominantly white?
Done using a transfer function Increases detail in the darker areas and lose detail in the bright areas. This is referred to as pulling down the colour scale Example function T(x) = x^3
What is the consequence of changing a grey scale to predominantly black?
Decreases the detail in the darker areas and increase the detail in the bright areas
What happens if you invert the colour scale used in greyscale visualisation?
Produces the same brightness and contrast but just has a different visualisation T(x) = 1-x
What is the principle behind colour coding visualisation?
Paint by numbers Have a defined minimum and maximum and you can choose the colours within the scale. More freedom Must display the colour scale next to the image
What is the advantage of colour coding?
Can bring out features not visible in black and white More freedom Many different colour visualisations possible Keeping the same colour for all images of the same type can allow for easy identification of pathology e.g. green - cancer
How does colour coding work?
The centre of the pixel is the colour determined from the relevant look up table Use a gradual transition to neighbouring pixels using an algorithm This makes the image easier to visualise - does not add any extra information
What is a surface plot?
It allows you visualise the image as a representation of height Height is proportional to the value of the pixel It can appear as a wire mesh image or as a rendered image - gives a 3D appearance
What is a contour plot?
It looks for areas that have the same plot. It represents the landscape. If the lines are closer together then there is a steeper change in pixel value Can used filled contours to smooth images
What is a profile?
Shows the change in height Need to take 1 line across the image and you can then see a change of pixel values
What is fusion and what are the various methods for doing this?
used when you want to view 2 image together - multislice display - overlay - blending - checkerboard - screen-door fusion
What is multi slice display (fusion)?
place multiple images together Very simple but hard to match up zones from the 2 images
What is overlaying (fusion)?
cut out a section of one image and place it over the top. it masks the old image but it gives more anatomical information
What is blending (fusion)?
Allows you to interactively move through the 2 image. Blends between them. you have to keep switching between them
What is checkerboard (fusion)?
confusing radiologically It is an alternate checkerboard to each image Usefully scientifically but not diagnostically If you do it for each pixel, eye can’t distinguish the individual squares
What is screen-door fusion?
Checkerboard image with each square equal to one pixel The eye blends the same as the pixels are too small to be distinguished individually It is the most common way to fuse Applied in PET/CT
What are the methods of 3D visualisation?
Slicing Multislice display Animation display Triplanar display Projections Surface rendering Volume rendering Reformation Vector fields
What is slicing (3D visualisation)?
Take a slice of pixels and view Can be done in 3 plane although it is possible to make other orientations through reconstruction CT always measured in axial Reconstructions are lower quality and blurrier - this is because it requires linear interpolation to find the centre value It is best to view it in the orientation measured
What is multislice display (3D visualisation)?
all the slices are viewed next to each other
What is animation display (3D visualisation)?
Animate the slice together in a sequence
What is triplanar display (3D visualisation)?
It allows each plane to be seen at once, with each representing the others.
What are projections (3D visualisation)?
Ideally want to visualise the whole 3D image. Each line of that runs perpendicular to the viewing plane will pass through many structures and they can’t all be represented at once Can use: maximum, minimum or mean
What is the mean intensity projection?
3D visualisation as a 2D image The average intensity through each pixel line is taken Not very useful
What is a MIP>
Maximum intensity projection Most common projection Takes the maximum value from each line for pixels 3D data viewed as 2D image It will pick up the brighter structures in the body Useful in angiography - can see the contrast agent within the blood vessels
What are the problems with projection images
Want them to be thin to avoid too much projected on top of each other.
Why are thinner slices better for MIP?
Avoid too many things projected on top of each other Allows you to see more detail by removing the extra information from the overlay
What is surface rendering?
Surface extraction to visualise a structure Need to define the structure and outline the surface you want to view (done via segmentation and converting to binary data)
How does surface rendering work?
Uses an algorithm that mimics what happens if you shone a light on the surface Tracks the rays hitting the surface
When is surface rendering useful?
Cardiac imaging using CT Virtual CT colonoscopy
How a virtual colonoscopy done?
Bowels are filled full of air Then do a 3D CT image You can then do surface rendering from the inside to create an endosurface
What is volume rendering?
Track what happens to light as it hits an object For every voxel along the ray you assign an opacity value and a colour The opacity value assigns transparencies to certain parts, allowing you too see the whole 3D structure
What are the pros and cons of volume rendering?
CT dose is the same Complex technique What you see depends on visualisation used Incorrect visualisation can lead to misdiagnosis Computationally expensive Requires high calculation power Time consuming High computer storage space better than mip as you can see what structures lay in front or behind the structure
What is reformation (3D visualisation)?
It distorts the data to allow you to see something better It changes the data
What are the different types of reformation?
Curved reformation Mercator mapping Registration to anatomical template
What is curved reformation?
Involves using a centre line and stretching it out to get a flat representation Can be done for arteries or whole spine Allows for easier measurement of change in diameter
What is Mercator mapping?
Virtual dissection and unrolling by mercator projection Tube is around the colon From the centre line everything is projected onto the tube Cut and open up the tube to see a 360 time in 1 flat image Decreases reading time and doesn’t detriment the detection rate
What goes registration to an anatomical template involve (reformation)?
It allows you to see if a structure has an abnormal shape It uses a library of normal structures Measure the patient use complex mathematics to determine how much deformation is required to fit to normal This is then colour mapped e.g. red - high deformation
What are vector fields (3D visualisation)?
Scalar components are used in diffusion imaging The colours change if there is a change in brightness in the vector field Brightness = magnitude Colour = direction Colours are allocated to a fibre direction - black = no orientation Commonly used in white matter Each pixel has 3 values - unit length lines show direction (instead of viewing directional arrows can view streamlines)
What are the 2 ways that non-convolution filters work>
Fixed Block Sliding Block
What is a fixed block filter>
The image is broken down into a number of equally sizes tiles and a calculation is made on the number of pixels in the tole. All the pixels under the tile are given the resultant value
What is a sliding block filter?
A block is place in the top left of the image and a calculation is made on the value of the underlying pixels. In the processed image, the single pixel at the centre of the block is given the resultant value The block is then advanced one pixel to the right and the operation repeated for the rest of the image
What are the consequences of using a fixed block filter?
Removes a lot of the noise Very quick and simple The effect is not very helpful Averaging
What are ordinal (rank) filters?
They work by sorting the numbers in a sling or fixed block window into numerical order. One of the sorted values is then chosen as the value of the filter The centre value is then chosen as the maximum, minimum or median value. Much more time consuming
What is a median filter?
Takes all of the values within a block, places them in numerical order. Calculates the median and places this value is the centre of the block. Under every pixel the median is calculated and each pixel is given this value. Reduces noise - replaces a pixel by its local average. Using the mean, less edge information is lost.
What is the consequence of using a median filter?
Often used to reduce noise Retains much of the structure Removes noise If done with an 8 pixel radius filter = 16x16 pixel blocks
How do you define convolution of 2 continuous 1D function?
f(x) x g(x) One function is the input image you are dealing with g(x) is the processing step performed on the image The output is the result of the convolution and the processing step. Multiply the 2 steps together and then integrate the result You will only get a signal when the 2 functions overlap
What are the steps in convoluting an image?
- Kernel placed in the first available position 2. Multiply each pixel by the kernel value 3. Sum up the values 4. Place the value in the centre pixel hotspot 5. Advanced to the next position 6. Filling in the edges
What are the methods use for edge filling?
It is standard practice to trim the output image to the same size as the original - Zero padding - Can just choose to have resultant image 2 rows and columns smaller than the original - Repeat or extrapolate edge pixel values - Mirror image to increase the size
What are the steps in convolution?
Reverse and shift the input function Pass it through the filter As it goes through the filter it is multiplied The integral of the result is then calculated Get the output You will only get a signal when the 2 functions overlap
What is deconvolution?
It is the process of figuring out what the filter has done Know input and output
What is a sharpening filter?
Negative Value on the left and right side Positive Value in the centre
What is the kernel?
It the filter that is applied to the image
What is zero padding?
Adding an extra row/column of zeros to the outside of the image It allows you to apply the kernel to all pixels in the original image The zeros have no impact as the filter involves multiplication The output image
What are the disadvantages to zero padding?
If there is an area of strong signal along the outside of the image and there is a zero next to it, it can create an edge in the image.
What is the most common edge filling technique
Zero padding or repeating edge pixel values
Why would you choose to blur an image?
Can be used to reduce the effects of noise If further analysis e.g. segmentation is required, blurring can make this easier Degrades the spatial resolution
What is the simplest method of performing blurring?
Local area averaging Box car blur
What is a box car blur?
It is used to blur an image using local area averaging For an area n x n pixels, each kernel pixel is 1/n^2 e.g. 1/9 to calculate the local mean No scaling to the image No change to lighter or darker Need an odd number of rows/columns
What happens are you increase the size of the boxcar kernel?
Increase the amount of blur Reduce the spatial resolution Reduce the noise by a greater amount
What does a boxcar blur appear like on imageJ
1 1 1 1 1 1 1 1 1 Process - Convolve
What is a Gaussian blur function
Higher values in the centre of the kernel Smaller values on the outside of the kernel Emphasising the centre pixel value Trying to retain the information that is in the image By including some information from neighbouring pixels you reduce the noise in the image
What is sharpening?
Sharpening an image involves enhancing the difference between the current pixel and the local area Increases the visibility of the edges Increases the high frequency components within the image
What does a sharpening kernel look like?
Combination of a blurring and a scaling kernel High centre voxel Subtract a local average from that Bright central voxel and subtracting the average side values -1 -1 -1 -1 24 -1 -1 -1 -1
What happens if you increase the intensity of the sharpening filter?
Increasing the difference between the centre value and the surrounding values The larger the difference - the more moderate the sharpening effect is If you have a high value in the centre, you are just replicating the original image. If the centre value is lower, it produces a more dramatic sharpening effect.
Why is edge detection useful?
Useful when you are trying to find structures in a semi-automatic way helps you identify a structure of interest
What are the edge detection filter types?
Prewitt Sobel
What is a Prewitt filter used for?
Edge detection Vertical edges -1 0 1 -1 0 1 -1 0 1 Horizontal edges -1 -1 -1 0 0 0 1 1 1
What does a Prewitt vertical edge filter look like?
Vertical edges -1 0 1 -1 0 1 -1 0 1 Negative on the left, positive on the right
How does a vertical edge filter work?
Increasing the values of the bright voxels on the right hand side Decreasing the values of voxels on the left hand size Negative values on the left, positive on the right Increasing edges are bright, decreasing edges are dark
What does a Sobel filter look like?
Vertical -1 0 1 -2 0 2 -1 0 1 Horizontal -1 -2 -1 0 0 0 1 2 1
How do you remove increasing and decreasing edge information from an edge detection filter?
Use the absolute values from the horizontal and vertical edge images and add them together identifies any edges in any direction
What is the result of applying a vertical edge detection filter?
Lose almost all horizontal edges but keep all vertical data
What happens if the sum of the elements in the filter is not 1?
If you don’t normalise the filter, you end up scaling the image or decreasing the image. If after filtering you want to analyse quantitatively, it will have an effect on your data
How suitable is an 8 bit unsigned integer image for applying convolution?
0-255 Can’t deal with negative numbers so you can’t apply a filter with a negative value in the kernel e.g. edge detection filter Can’t handle fractions - can’t apply a blurring kernel that involves finding the average Can’t deal with large value so you will struggle to scale up Poor
How suitable is a 16 bit signed integer for applying convolution?
Can deal with negative numbers - can use filters with negative values Can’t scale down or find a local average as it can’t deal with fractions Could scale the image up and then apply the filter, however, unless the answers are whole numbers they can’t be represented. You can choose to select the nearest whole value in order to display it Can deal with larger values.
How suitable is a 32 bit floating point image for applying convolution?
Increased storage If the images need to go back onto the PACs system then they need to be done in a format that can be displayed on PACS Not necessarily able with this type Can deal with large values, negative value and fractions. Process of convolution should not be heavily restricted. Relatively large degree of accuracy
In convolution,the filter is flipped around the origin prior to application. When would this not have an appreciable effect?
If it is an edge detection filter (looking for a vertical line), if you do not flip the filter you will find rising instead of decreasing edges and VV
What is the consequence of using a maximum filter?
You get circular effects The maximum filter in the kernel is likely to be the same value for many positions of the filter and as a result you get bright patches which are circular. Strange spotty appearance Not useful in medical imaging Can be useful in microscopy
What is a fourier transform?
A mathematical tool that converts an image in the spatial domain to the Fourier domain (spatial frequency domain). Allows you to get frequency information out of an image It is possible to filter in the fourier domain and then do an inverse fourier transform
What are the properties of a Fourier transform?
It is a complex function If x represents time then u represents frequency (Hz) If x represents space then u represents spatial frequency (cycles/mm) Output is in complex number form
Describe an FT image in terms of the spatial frequencies?
General contrast = low spatial frequency Detail = high spatial frequencies
What are the advantages of using a FT?
Can be dealt with rapidly by computer programs Processed very quickly Useful in image processing - convolution of the image in the FT is simpler. It is just the multiplication of functions. No convolution step - much quicker
What does a FT do?
It extracts all of the information in a signal in terms of its frequency components. What sinusoids make up that signal - any signal that can be measured can be produced using a combination of sinusoids Interested in the phase and amplitude
What is a power spectrum?
It gives the energy within a signal Centred on the zero function Y axis is on a log scale
What does the imaginary number mean in FT?
Represents the phase information Allows you to represent complex signals
What is the phase in FT?
The phase difference in the sine and cosine sinusoid components of the image
How does a FT image appear?
Low frequencies at the centre of the image Detail (high frequency) is in the periphery of the image
What does a boxcar filter look like in the spatial domain?
Spatial filter Fixed flat response to a fixed value
What does a Gaussian filter look like in spatial domain?
Right half of a bell curve from the x=0
What does a sharpening filter look like in FT domain?
It has a high point at x=0, a rapid drop to a negative value and slowly returns to y=0
What does a low pass filter look like in the FT domain?
What does a low pass filter look like in the spatial domain?
What does a high pass filter look like in the FT domain?
What does a high pass filter look like in the spatial domain?
What is a low pass filter?
It only allows low frequencies through and removes any high frequencies
It removes all sharpe edges and detail
Keep the centre of the FT only
What is a high pass filter?
It allows high frequencies through but not low frequencies
Looks like a sharpening filter
Remove the centre of the FT
Retain the high frequency information but can get a ringing effect due to the response of the filter
1- low pass
What is an intermediate filter?
It will only let through a narrow range/band of frequencies (Band pass)
Very complicated in the spatial domain and it not typically used in the raw image. Nearly always done in the FT.
What does an intermediate filter look like in FT domain?
What does an intermediate filter look like in the spatial domain?
Not normally done in spatial domain
Complex
What is the impulse response of the filter?
Ideally you would want a single value but it spreads out from the centre
The larger the spread the larger the impulse response
How can the rippling effect of a high pass filter be removed?
It can be mitigated by using filters that have a smoother transition between the inside and the outside of the pass band
This comes at the expense of extra high frequency information/detail.
The smoother the value between large and small values, the less artefacts there are.
What are the different filters for frequency response in the FT?
Gaussian filter - you set the width of the filter by altering the standard deviation
Butterworth - similar to Gaussian but has a tighter frequency response
Hamming
What is the purpose of the filters applied in the FT with high or low pass filters?
They try to acheive a small range of frequences through the filter while reducing the artefacts generated
How can periodic noise be removed from a FT?
Spots can be seen in the FT and then remove those 2 points using a band stop filter
It is very specifc
How do you decide where/how to apply your filter? Spatial or FT?
It depends on which will be more efficient
FT can simplify something complex in spatial
Artefacts can be removed more readily in FT
Ease of representation - things are often seen easier in spatial domain
What is the Hough Transform?
It is a method used for detecting shapes in images
Most commonly straight lines or circles
It is computationally fast and copes will with noisy images
Insensitive to noise
Can cope with missing gaps or if not perfectly straight
What are the steps used in a Hough transform?
- use edge detection filter on the image
- Threshold thie image to create a binary image
- Perform Hough transform on the binary image
- Locate the first non-zero pixel in the iamge
- Cycle through all possible points in paramaeter space whose corresponding line passes through the pixel.
- Add one vote to the point at each time
- Repeat for all non-zero pixels - Find peaks in the Hough Transform - Brightness depends on the number of votes
- Identify the lines
What are the difficulties in finding lines within an image?
A line has a gradient and a y-intercept
The problem each the numbers for these are infinites. Any one point can have many lines passing through it.
For every point: the line is desribed as having an angle and a distance.
What is image registration?
The determination of a geometrical transformation that aligns the position of feature in one image of a subject with the position of the corresponsing features in another image of that subject or another subject
What are the 2 categories for medical imaging applications?
Anatomical - depict morphology
Functional - depict information on the physiology
When can registration be performed?
2D to 2D
3D to 3D
2D to 3D
What are the processes involved in 2D to 2D image registration?
Does not require scaling
Only rotation and translation
If the geometry of the image acquisition is rightly controlled then they can be registered simply.
e.g. NM and X-ray alignment
What are the applications of image registration?
Aid to diagnosis, surgery and therapy
Diagnosis can be aided by combining functional and anatomical data
Planning of radiotherapy and surgery can be aided by combining
Response to therapy can be assessed by measuring changes in serial studies
What are the different types of image registation?
Intramodality intersubject
Intermodality intrasubject
Intermodality intersubject
What does intramodality intrasubject imaging involve?
Alignment of the same subject in the same modality
Used for serial studies taken over a period of time
Example: assess disease progression, monitoring the contrast agent uptake over time
By bringing images into spatial alignment, can measure any changes. Allow for statistical or matheamtical operations
Why is image registration important?
You need the pixels in the same position in before and after studies for time activity curves to be accurate
What can cause misregistration and how can it be detected?
Can be causes by incorrectly aligning the images together
Caused by patient movement
It will be seen on a subtraction image - if the edges are prominnet then there is a misregistration
What does intermodality intrasubject image registration involve?
Alignment of the same subject in different modalities. It is used to combine different informaation
e.g. PET/CT
Needs scaling and rotation for alignment
Example: Radiotherapy treatment - aligning CT with MR (MR for delineation of tumours, CT for densitiy information)
What does intra-modality inter subject registration involve?
Alignment of different subjects using the same modality.
The objective is to statistically model the anatomy of an organ across subjects - useful when group comparison is needed and is used to build atlases.
Can study the variability between patients
What are the different types of transformation?
Rigid body
Affine
Non-affine
What is a rigid body transformation?
6 degrees of freedom - 3 rotation, 3 translation (x y z for each).
NO SCALING
All distances are preserved
Special form of affine (skew to 0, scaling set to 1)
When is a rigid body transformation carried out?
Used for bone, structures encases in bone (brain) and structures in the vicinity of bone e.g. neck and spine
Simple - when something can’t deform
When all parts of the object are assumed to move as a whole
What is an affine transformation?
12 degrees of freedom - 3 translation, 3 rotation, 3 scaling values, 3 skews
Rigid body is a Special case of affine
All lines remain straight and parallel lines are preserved.
When is an affine transformation used?
Used for scanner induced error e.g. skew in CT
Angles can change
What is a non-affine transformation?
Many more degrees of freedom
Described by the deformation field with a displacement at each voxel
Most organs in the body do not just stretch and shear
When is a non-affine transformation used?
Used for deformable organs in the body e.g. liver, bladder, heart, prostate
What are the models used for non-rigid registration?
Done using deformable transformation models:
- B spline and thing plate spine (common-used for parametrised transformation fields)
- Non-parametric: There is a displacement vector at every grid location in image B needed to align with A
Good for breast imaging
How do you quantify the registration quality?
Similarity measures
- Landmark, surface or edge measures (sum of squared differences, correlation coefficient)
- Voxel similarity measures (mutual information)
Describe the image registration algorithm
Start with the source image B
Select initial transformation parameters
Transform
Evalulate similarity
Is the image best matched?
NO = update trasnformation parameters (iterative loop)
YES = Registered image
What are the landmark measures used in image registration?
Anatomical landmarks
Extrinsic markers
Use these to calculate the sum of squared distances - want this to be as small as possible
What is required to use anatomical landmarks in image registration? What are the pros and cons?
High resolution images
Difficult to autodetect
Manual
What are the pros and cons of using extrinsic markers for landmark measures?
Impractical for routine use
Well suited for validation studies
Easy to autodetect
(Stereotactic frame = can be seen in MR and CT using contrast agent)
What are surface or edge measures for image registration?
Techniques based on surface information using: points on surface or characteristics of a surface edge e.g. crest lines
Correlation coefficients can be used a similarity measure
What is the algorithm used in surface or edge measures in image registration? How does it work?
Hat and head algorithm
- Transform the hat surface (iteratively) with respect to the head
- Until you find the closest fit
- Measure the distance between point on hat and nearest point on head in the direction of the centroid of the head
- Minimise this distance
What are the applications of the hat and head algorithm?
Skin or brain surfaces
PET MR
What is a distance transform?
Can increase the performance of hat and head algorithm
It is applied to a binary image and each voxel is labelled with the distance from the surface of the object
It simplifies the iamge to make it quicker and reduce the computational cost
What is the most common distance transform?
Chamfer filter
What is Chamfer matching? What are the steps involved?
A modification of a technique by pre-processing head images to be used for hat head algorithm
- Generate a binary image
- Apply a distance transform (Chamfer filter) to generate an image when pixels are labelled with their distance from the surface
- Iteratively determine T
What does a Chamfer fitler do?
Generates an image where the pixels are labelled with their distance from the surface
Simplifies and speeds up the process.
Why use a Chamfer filter?
Simplifies and speeds up the process
Registration using distance maps should be more robust
What are voxel similarity measures?
Assume correlation between groups of voxels that have similar intensity values. The measures are calculated directly from voxel values:
- Entropy
- Joint entropy
- Mutual information
What is entropy in relation to image registration?
Image registration is trying to maximise the amount of shared information in the 2 images or minimiseing the amount of information in a combined image
Entropy is a statistical measure of randomness that can be thought of as the measure of information in an image
What does high entropy indicate?
An increase in entropy indicates an increase in the amount of information.
Poor registration = high entropy in the subtraction image
How can entropy be approximated?
Using the image histogram
What is joint entropy?
Measures the amount of information in the 2 images combined - can be visualised using a joint histogram
It disperses or blurs with increasing misregistration
Describe a joint entropy histogram
If both images are of the same modality then there will be a straight line at 45 degrees through zero if perfectly matched, this line blurs with increasing misregistration
If mixed modality - it does not produce a straightline as the images look different but you can still aim to minimise the entropy in the image
What is mutual information?
It is a measure of how well one image explains the other.
Maximising the mutual infromation is equivalent to minimising the joint entropy
Considers the intensity distribution of both image data sets
What is the advantage of using mutual information over joint entropy?
Mutual information includes the individual images entropy
Works better in regions if image background (low contrast) where there will be low joint entropy but this is offset by low individual entropies as well so the overall mutual information will be low
Where is mutual information most well suited in image registration?
Multimodality image sequences as it makes no presumptions about the intensities
e.g. it is possible that ergions with a high intensity in the reference image coorespond to regions with a low, medium or hgih intensity in the source image
When does mutual information reach a maximum
For a given intensity in image A, a distinct intensity in image B is found
What are the practical considerations for image registration?
Pre-processing
Interpolation
Improving speed
What are the pre-processing considerations that need to be made in image registration?
Segmentation - identify the structures
Exclusion of non-fixed anatomy e.g. femur
Binary image and filters e.g. Chamfer map
WHen is interpolation used in image registration?
Required to estimate the values of the transformed image e.g. if the image is rescaled to be larger
Can lead to smoothing of the image and involves finding a best estimate of the value of a point between pixels
How can you improve the speed of image registration?
Sub sampling
- Image volume during early iterations (cut down number of pixels sampled in early interations)
- Intensity histogram
How can qualitatively assess image registration?
Subtraction
Visualisation
- Dual cursor
- Checkerboard
- Overlay
- Surface displays or cut surface display
How can subtraction be used to assess the quality of registration?
Estimates registration accuracy by inspecting the images
Quick method = subtraction
If images are perfectly aligned then they will have zero pixels on subtraction image and the image would appear mid grey. Areas of misalignment are much darker or brighter
Why is visualisation useful in qualitative assessment of image registration?
You can pick out features and check alignment
Good for checking edges
Quick
(overlay, dual cursor, checkerboard, surface displays)
How is image registration assessed quantitatively?
Evaluation of the registration method to a gold standard ir refernce method
When developing a new technique it may mean comparison against a manual technique
The problem is that the reference method may not be that good and this can be challenging
Can get an error value in registration which is particuarly useful in radiotherapy
What is radiotherapy?
Use of high-energy radiation 6MV-15MV to treat cancer. The X-rays interact with patient cells causing damage.
Tumour is accurately delineated to prevent damage to healthy tissue.
50% receive radiotherapy as part of their treatment
What are the steps involved in radiotherapy?
- Diagnosis and decision to treat
- Immobilsation
- CT Scan
- Delineation
- Treatment planning
- Treatment delivery
Why does a patient need to be immobilised in radiotherapy?
Radiotherapy is given in a discrete number of fractions. This allows for the recovery of healthy tissue (up to 40fractions over 8 weeks)
Patient position needs to be reproducible so they are in the same position for treatment as they were in the planning scan.
How are patients immobilised for radiotherapy?
Minimise the degree of movement and the increase the reproducibility of patient set up
- Masks for head and neck
- Rests to lift the arms out of the way
- Leg rests to hold the hips in the same position
Delineation: what happens in radiotherapy?
Ct scan is assessed by a clinician
Tumour target delineated on the dataset in addition to any radiosensative organs that must be avoided - Do not want to over radiate OAR e.g. spinal cord
How is radiotherapy treatment planning done?
Computer software is used to positon patient beams upon the marked structures
Resultant doses are calculated
It is an iterative process in which many parameters are optimised until a clinically acceptable outcome is reached
Combination of beams from different directiosn in order to achieve a high dose at the tumour
How is radiotherapy treatment delivered?
High energy radiation (6MV to 15MV) produced using a linear accelerator
Accelerate electron to tungsten target, patient is positioned under the beam
The beam of radiation varies in sizes and is delivered at different angles based upon treatment plan.
Lead collimators allow you to change the shape of the beam
What are the 2 key components to successful radiotherapy treatment? How is at achieved?
Full target coverage to prescription dose
Minimial dose to surrounding healthy tissue and organs at risk
Achieved by complex planning techniques and a highly conformal dose distribution (volumetric modulated arc therapy)
What is the risk of having steep dose gradients in radiotherapy?
Highly conformal dose distributions but an increased risk in a geographical miss
Loss os local tumour control and effectiveness of treatment
Higher risk of irradiating OARs
What is the patient pathway in radiotherapy?
Diagnosis
CT scan
Target and critical organ definition
Treatment planning
Treatment delivery
Verification to treatment planning
Follow up
Where is the patient radiotherapy pathway is imaging used?
Target and critical organ delineation
Verification
What happens during delineation of tumours?
Outline the target region to which the high dose should be delivered
Also outline OARs
Require a CT for dose calculation purposes
Contours are created in each 2D slice to see it in 3D
What are the probelms associated with using different imaging modalities in radiotherapy planning?
The patient may be in a different position
It may have been taken at a different time
In a different radiographic plane
Produced in a different format e.g. film/radiographic
What are the image registration techniques used in radiotherapy?
Lanmark based - user identifies landmarks in each modality and software aims to minimise the error between the 2 positions
Surface based - surfaces in each are designed and mutally regisitered (works well for bone)
Drag and drop - can interactively rotate and translate one data set over the other (manual)
Automated - mutual information and least entropy methods
What are the visualisation tools used in radiotherapy?
Checkerboard
Horizontal or vertical slider
Spyglass
Rectangular window
Blend/fusion
What is image verification in radiotherapy?
Process of ensuring the actual beam placement matches the planned one and you need to quantify the magnitude and direction of the error
How is image verification done in radiotherapy?
2D; compare MV planar imaging against digitially reconstructed radiograph
3D: compare 3D cone beam CT with treatment planning CT
What is the EPID?
electric portal imager
It has an amorphous silicon flat planel detector and it allows a 2D digital image projection to be acquired using the MV treatment beam
How is MV planar imaging used in radiography?
To assess the setup error
The MV images are compared to digitally reconstructed radiographs created from the CT planning scan
The image quality is poor - can only be used for bone
Need AP and lateral projection
AP can calculate left right and sup/inf error and lateral shows ant/post
Why do anatomical changes frequently occur in radiotherapy?
Often several weeks between the planning scan and treatment
Fractionation schedule can be up to 8 weeks
Concurrent chemo patients are most susceptible
What is the benefit of using 3D imaging to check patient position in radiotherapy?
You can image the patient in the treatment position
You can then perform image registration with the planning CT and quantify and positional errors in patient set up
It has adequate soft tissue contrast so you can determine if any weight loss has taken place or if there has been any soft tissue deformation
What are some of the functional properties that can be obtained from imaging?
Perfusion
Permeability
Diffusion
Flow
Velocity
Molecular imaging
Oxygenation
Spectroscopy
Tissue composition
Elastography
Temperatire
What is the advantage of using imaging biomarkers?
It can increase the personalised aspect of medicine
Imaging is moving from morphological to multi-parametric imaging
Define quantitative
a type of information based in quantities or else quantifiable data and gives objective properties
Define qualitative
Information which deals with apparent or subjective properties
What imaging modality is quantitative?
Unenhanced CT
Different tissues have different mass attenuation coefficients annd so this provides contrast
Describe the units used in CT
Hounsfield units (HU)
Relative to water
Narrow range for soft tissue and this is why soft tissue is poor
u x - u water / u water - u air x 1000
Why is unenhanced CT qualititative?
Each pixel intensity tells you something.
You can determine abnormalities from the value
Describe CT windowing
It is imprtant and needs to be correct depending on the structures you are interested in looking at. It is different for different structures.
Bone window = high centre, wide width
Soft tissue window = low centre, narrow width
Lung window = very low centre, wide width
Give an example of quantification using CT
Haemochromatosis - increased HU in pathology due to increased iron deposition (high u, high HU)
Adrenal adenoma - adenomas are fatty and so have a negative HU and appear darker.
Staging of fatty liver - increasing stage = increase fat = decreasing HU value
What are the features of quantitative imaging?
Can be reproduced on another scanner
Both contrast and intensity are meaningful
Pixel values are a measure of a well-defined physical or physiological tissue characteristic
Why is contrast CT not quantitative?
Constrast stops this
The quantity of contrast, time to acquire the image, time for the bolus injection can all change the intensity
Need to remove to get qualitative data
Can do semi-quantitative - no measurements but can be a good indicator
How do you quantify contrast enhanced CT?
- Contrast concentration. Need to convert the change in HU to a concentration in the plasma, The can be done using calibration curves
- Arterial phase image - non enhanced image = concentration - Functional parameters - calculate extracellular volume fraction or other parameter
What are the complications of using calculating contrast concentration in CT?
Need a pre contrast image to subtract and as a result the dose is higher
You then need to match pre and post contrast.
If there is movement this can corrupput the data and produce bright rings
How do you calculate extracellular volume fraction?
Tissue concentration (mg/ml) / Plasma concentration (mg/ml)
Measure plasma concentration
Get tissue concentration from the image
Why is calculating extracellular volume fraction useful?
Contrast leaks out of the blood into extracellular space
This can indicate if a tumour is benign or malignant
You can then map it and it gives quantitative physiological data
It is very process heavy
What are the parameters that can be calculated using enhanced CT?
Extracellular volume
Blood volume
Bloow flow
Mean transit time
Capillary permeability
Interstitial volume
Give a use of functional CT
Acute stroke
Assess if tissue is viable before reperfusion e.f. if it is over 20 ml/min/100g then it is reversible
Monitor response to treatment and help with prognosis e.g. tumour therapy with chemotherapy drugs
What are the advantages and disadvantages of quantitative unehanced CT and enhanced CT?
Unenhanced:
Cost - none, benefit of attenuation values, standard application
Enhanced
Cost: dose, analysis time, image quality. Benefit of physiological parameters. Application is emerging
What are the requirements of a computer aided diagnosis in CT colonscopy?
Needs to find lesions, define it and classify it
Aims to remove the human element.
What are the current applications of computer assisted diagnosis? What is the purpose?
Assist the radiologist
CT colonoscopy
Mammograms
Lung imaging
Detection of vertebral fractures - very subtle, easy to miss.
When is CAD useful?
Screening: large quantities of patients, most wilth no or small lesions.
Computer aims to pick up any suspicious areas that can then be assessed by a radiologist
e.g. mammography
Why is CAD useful in the lungs?
Finds nodules and can determine from previous scans if they have changed in size and number
In the lungs there are many slices and it canbe hard to determine what is suspicious and it can be very easy to miss nodules.
It aims to flag any suspicious nodules for further assessment
What is the purpose of computer aided diagnosis?
Detect a number of structures
Classify each as positive or negative
Present positive structures to the user for inspection
It reduces the amount of work to be done
What are the outcomes of CAD?
True positive: correctly classified as positive
False positive: Classified as positive when actually negative
True negative: correctly classidied as negative
False negative: classified negative when actually positive
Define area
Area of selection in square pixels or in calibrated sqaure units (mm2)
Number of pixels (find on histogram) x pixel size
Define standard deviation
Standard deviation of the values within the ROI
It is a measure of the width of the histogram
Define mean grey value
Average grey value within the selection
Sum of all of the grey values / number of pixels
If colour r + g + b /3
Define modal grey value
Most frequently occuring grey value and corresponds to the highest peak in the histogram
Define centroid
The centre point of the selection
Average of the x and y coordinates
Uses X and Y headings
Define centre of mass
The brightness-weighted average of the x and y coordinates in the selection
Uses XM and XY headings
First order spatial moments
Define perimeter
The length of the outside boundary
Define boundary rectangle
The smallest rectangle enclosing the ROI
Gives width and height as well as the coordinates of the upper left corner of the rectangle
Define Feret’s diameter
The longest distance between any two points along the selection - also known as maximum caliper
Define integrated density
The sum of the values of all pixels in the image or selection
Equal to the product of the area and mean grey value
Define skewness
Measure of how asymmetric a histogram is
Third central moment of the histogram, normalised with the cube of the standard deviation
What does thresholding do and how is it useful?
Creates a ROI by selecting only pixels in a given intensity range
Quickly and semi-automatically segment out particular anatomical structures - provided the image has structures with distinct intensities
How does an edge detection filter work?
Identifies the boundaries between structures
Can determine the gradient of the image and then apply thresholding only selecting the brighter pixels
This will include boundaries as there is a steep change in values
How does a contour plotter define an ROI?
It creates contours at given user defined values
Contours will often be closed lines which means they can be used to define a ROI consisting of all pixels within the contour.
If there are multiple contours it will select the one with the largest area for ROI
What happens when you downscale an image and then upscale? x4
Information is lost and can’t be recovered
Each 4x4 grid initially all had different pixel values but now they all have the same value
Decrease resolution and decrease detail
What effect does zooming have?
Uses interpolation to make the image bigger
It does not change the actual data stored
Just visual
What is the effect of averaging when downsizing?
If you scale down by a factor 4 with averaging then each 4x4 sqaure is replaced with a single pixel value equal to the average of the original 4x4 area
Without averaging, one pixel value is selected
Creating an average of N values decreases the standard deviation by root N
Reduces noise level
Define brightness
Centre of the greyscale
Average of maximum and minimum
Max + min/2
Define contrast
The range of the scale
Max-min
Window - width of greyscale
Why does increasing kernal size increase blurring when smoothing?
If the pixel has radius 20 then the calculated vlaue of a given pixel is dependent on all pixels in a 20 pixel radius
How depends on filter choice (mean, gaussian,median)
Radius of 20 = decrease in resolution by factor 40
Smaller radius = smaller reduction in resolution
What is a profile? How is it useful?
Plot of signal intensities along a given line
Useful to quantify signal change
Numbers aren’t shown on the image but can be read off the pofiile
Can visualise changes in intensity gradients ro difference in intensity in different areas
Appearance is independent of grey levels or colour settings
What is screen door fusion? How is it useful?
View 2 images at once
Creates a new image by alternating pixels from both
Checkerboard with sqaure size 1 so eyes can’t distinguid
Looks like superimposed transparency 50%
What is the effect of using a median filter on file size?
The output image is the same size as the original and so the file size will be the same
Add zeros, repeat out edges or mirror to get the same size
How can you tell what edge filling technique is used?
If it is smaller will have less pixels.
If zero padding then use edge detector and if there is a bright pixel on the edge it will create an edge
If mirroring, then if there are bright values along top and bottom then there may be bright values outside expected area
If repeating - ImageJ
What is a low pass filter?
Keep low
Reject high
Clear the area outside the circle
All low frequency information - contrast
What is a high pass filter?
Keep high values
Remove low
Clear inside the circle
Keep high frequency = detail
Why are CT images useful in planning treatments?
Provide density information for dose calculation
Provides geometrical accuracy for the beam setup
Why would you use MR in radiotherapy treatment planning?
Improve target colume definition
This is due to high soft tissue contrast and it is easier to define the tumours and OAR
Why may the prostate by misregistered even if the chosen structures are aligned perfectly?
Organ motion may be due to:
Different degrees of bladder filling or
Peristalsis
What are the disadvantages of CT-MR manual registration of prostate?
2D not 3D
Operator dependent
Time consuming
Anatomical landmarks can be hard to identify and may be limited number in the image
Poor reproducibility
Why would a histogram based segmentation method not be useful for segmetinting bone in MR?
In MRI bone does not have a signal intensity associated with it.
This may be due to motion, fold over and chemical shift
The lack of protons in cortical bone means that it does not produce an MR signal
Why is Chamfer matching a good technique for automatic registration of the pelvis or MR and CT in radiotherapy?
It works well with poorly segmented (noisy) data seen in MR
It is suited to bony anatomy which is used in radiotehrapy treatment
Define specificity
Number of true negatives/ total negatives
Define sensitivity
Number of true positives/total positives
What can be said if a test has high specificity?
Positive result confirms the disease
What can be said if a test has high sensitivity?
A negative test will rule out the disease
What is the methodology of CAD?
Preprocessing
Selection of candidates
Feature extraction
Removal of false positive outliers
Classification
What preprocessing steps can be used in CAD?
Background subtraction
Selective enhancement filters e.g. dot enhanced or line enhanced
What are the features of candidate selection in CAD?
Pick up possible candidates
Make sure true positives are included
Very sensitive
Not specifc
What is feature extraction in CAD?
Need to define characteristics of lesion to describe what each looks like e.g. diameter
How are false positives removed in CAD?
Reduce the number of data points
Eliminate points that are definitely negative based on features
How are lesions classified using CAD?
It is complex
For each lesion there are many characteristics that can define it and there is often overlap
Collect training data and then measure the volumes - determine whether malignant using biopsy
Maximise sensitivity and specificity
