The raw data from a microarray experiment is a series of scanned images.

• Images must be converted into quantitative data
• Steps to preprocess and transform the image into a format suitable for analysis are under the realm of "image analysis."

• One array has two images

1. Cy3 dye, green channel, 510 - 550 nm
2. Cy5 dye, red channel, 630 - 660 nm

• These channels are distinguished by a scanning instrument
• Data for each channel are stored as monochromatic images

• Consider the computer image to be a two dimensional array (matrix) of numbers.
• The smallest element of the image is a pixel.
• For an image with M x N pixels, each pixel has location \((x,y)\).
• Each pixel has an intensity value \(f(x,y)\), and the size of the pixel is \(\Delta x*\Delta y\).

• For a monochromatic image, \(f(x,y)\) is an integer called a grayscale value where \(f = {f(x,y): x = 0, 1,..., M-1; y=0, 1,..., N-1}\).

• Therefore, each \(f(x,y)\) represents the brightness of a small picture element, called pixel, at location \((x,y)\).

• The number of pixels contained in a digital image is called resolution

• Pixel intensity values are stored as binary numbers.
• Binary numbers are sequences of 0's and 1's.
• A binary digit (usually abbreviated 'bit') can hold one binary digit, a 0 or a 1.

• For a one-bit digital image, the computer uses one bit to represent a pixel value. The pixel is either 0 (black) or 1 (white).
• In a two-bit digital image, the computer uses two bits to represent a pixel value. The pixel may be 00 (black), 11 (white), or one of two shades of gray (10 or 01).

• Most images are stored using n-bits; there are \(2^n\) possible binary sequences of length n.
• For example, an 8 bit (one byte) image has \(2^8\) possible values ranging from \(0\) to \(2^8-1\); \(0\) (black), \(255\) (white), and \(254\) gray levels.
• The radix is 2 so each number is represented as linear combination of powers of 2.

• Digital image: rectangular array of intensity values
• Each intensity value corresponds to a pixel
• Color Depth: is the number of bits used to store the intensity value of one pixel

Color depth of 16 bits/pixel (common for microarray scanners) means the intensity values of each pixel is an integer between \(0\) and \(65,535 (=2^{16}-1)\)

• Addressing: locate the spots
• Segmentation: categorize each spot as foreground (signal), background, or other
• Intensity extraction: assign signal and background values to each spot
• Spot quality assessment: compute measures of spot quality for each spot

These steps use specialized software and can involve varying degrees of human intervention.

• The ultimate goal of any image analysis technique should be the automation of the image analysis process.

• Although the layout of the cDNA array is known and can be used for addressing, the known model must be matched to the scanned image.

• Therefore, most software packages include both automatic and manual procedures for addressing.

• Once the address of the spots has been identified, the pixels must be classified as signal versus background, a process called segmentation.

• Background represents a value of the measured signal intensity that is presumed to be due to non-specific binding of target to the probe

• Thought to be removed from the signal intensity measurement in order to accurately quantitate the amount of target RNA present in the sample.

Uneven hybridization, auto fluorescence, non-specific binding - measurements outsize the spot not at 0 intensity

Spatial based segmentation:

• fixed circle
• adaptive circle
• adaptive shape

Intensity-based segmentation

• Ranked intensities
• Mann-Whitney method

• Get intensities of all pixels in a rectangluar region
• Estimate the number of pixels in a circle fitted in this region
• Select this number of intensities from the ranked list of all intensities

• A rectangular region, \(40x40\) pixels: a total of 1600 pixels
• Assume signal spot being 20 pixels in diameter: contains \(\pi *10^2 = 314\) pixels, or \(~20%\)
• Select \(20%\) top intensities from the ranked list of all 1600 intensities

• Chen et al (1997)

Chen Y, Dougherty ER, Bittner ML. "Ratio-based decisions and the quantitative analysis of cDNA microarray images." J Biomed Opt. 1997. PMID: 23014960

http://bcb.dfci.harvard.edu/~gp/teaching/688/chen1997.pdf

• Again, the purpose of segmentation is to partition pixels into one of two classes, foreground versus background.
• Most often, correcting for background takes on the form of subtracting the estimate for background from the estimate from signal.
• Subtracting background has been noted to increase the variability of genes, particularly at low levels of expression.

Spot intensity: Some statistics representing intensities for all pixels in spot area; similarly for background intensity

• Mean: mean of pixel intensities
• Median: median of pixel intensities
• Mode: location of peak in histrogram of intensities
• Area: number of pixels
• Total: sum of pixel intensities

Still, no consensus what to use

• The underlying principle that should be used to guide the selection of a method for data quantification is to select the statistical summary that best correlates with the amount of DNA target in the hybridized sample.
• L-estimators - linear combination of order statistics

• Affymetrix Genechips use propriety Affymetrix software
• Genechip surface covered with square shaped cells containing probes
• Probes are synthesized on the chip in precise locations
• Thus spot finding and image segmentation are not major issues