iCluster User Manual v1.03beta © 2009 Nick Hamilton
Contents
1. Statistical Test checkbox is clicked the user is presented with 2 lists of the image classes Those image classes selected in the left hand list are then tested against those selected in the right hand list Once classes have been selected click the Finished Selecting Classes checkbox and the test will start The test is for the null hypothesis that the image sets come from the same population more specifically that the mean vectors of the population that the two sets are sample from are the sample This is performed using permutation testing as described in 1 A p value for the null hypothesis is returned The number of repeats to give the p value may also be selected though under most circumstances the default of 1000 should be sufficient The Euclidean distance between the centroids average statistics vector of each cluster are also returned To exit the Statistical Testing mode click the Statistical Test checkbox Note Results may vary according to whether Normalise Statistics below and Select Statistics below are set Selecting Statistics The statistics that are used within iCluster may be selected or deselected by clicking on the Select Statistics checkbox and then checking unchecking the listed statistics It is probably safe to ignore this unless you want to use iCluster with your own image statistics rather than the automatically generated threshold adjacency statistics If you do use2. be viewed by pressing the m key while images are being viewed Press m again to return to viewing the images In cases where the automated masking fails completely the user may supply their own masks for each image A mask for an image is an image for which the non zero intensity pixels select the regions of interest in the original image See Sample Image Description File section below for how to supply a mask image for an image Snapshot Images Pressing p will take a snapshot image of the iCluster window The image is automatically saved to the directory containing the iCluster executable Sample Image Description Files There are a number of formats which will allow images classes statistics to be defined in an iCluster Image Description File Essentially the file is plain text with a single line for each image comma separated describing its name class es possibly a mask possibly user supplied statistics Hence exporting an Excel document with appropriate column entries to comma separated format text file csv should produce a correctly formatted file though the column headings may need to be removed In it s simplest form in each line an image name and a class is given for instance 010 _mitotracker png mitochondria defines the image 010 _mitotracker png to belong to the class mitochondria To supply a mask with each image a each line might look like 010 _mitotracker png 010 _mitotracker_mask png
3. is possible Click on Sammon Map Statistics to start the process It is iterative and so you will see the images fly around the screen a bit as they find their preferred location After a period of time the images should become relatively stable in their location at which point you may click the checkbox again to stop the process Note Results may vary according to whether Normalise Statistics below and Select Statistics above are set Normalising Statistics Statistics may be normalised that is for each statistic then mean of the statistic is subtracted and the result divided by the standard deviation for that statistic Threshold adjacency statistics are standardised for the complete image set when first calculated see below However if you wish to only compare some subset of the complete image set it may be appropriate to normalise the statistics with regard to that subset Hence when Normalise Statistics is checked normalisation is performed with regard to those classes of image that are visible i e their class checkboxes are selected Calculations such as representative image selection high dimensional nearest neighbours sammon mapping are then performed with respect to the normalised statistics for the visible images In the case of statistical testing normalisation occurs with regard to those classes of images that have been selected for comparison Threshold adjacency statistics are standardised when f
4. pressing while iCluster is running This will reduce the amount of memory needed for the visualisation but will mean that images are not seen in such detail iCluster has successfully viewed a 1000 images of individual cells simultaneously by using a good graphics card on a notebook computer and reducing the image resolution iCluster doesn t do anything when starting Try deleting the file config txt in the directory with the executable iCluster may be getting confused about where the default directory image set is Again large image sets may cause problems dues to memory limitations Windows Vista seems to be particularly inefficient in this respect The Linux distribution of iCluster will readily display twice the number of images that Vista will on identical hardware One work around for this is to reduce the image resolution displayed when iCluster starts In the folder containing the executable is the file config txt This file contains 3 lines of information the default directory of images the name of the image description file and a line that reads iScale n where n is an integer This is a scaling factor used to reduce the image size For instance if n 3 the images will be scaled to 1 3 of their original size Hence larger image sets may be displayed by increasing n Note that this only effect the display of the images statistics are still calculated on the raw images
5. Note for video memory reasons images are initially viewed at 1 3 of their size This resolution may be increased decreased using Increasing the resolution too much may lead to memory problems and the iCluster to crash for large image sets depending on how much video ram your computer has Simililarly displaying image masks see below doubles the amount of video ram used and may cause problems for large image sets at high resolution A recent dedicated graphic card is recommended for use with iCluster 2D 3D Representation Image sets may be viewed layed out in 2D or 3D 3D will typically be better in the sense that the Sammon mapping see below of statistics will better preserve the distance relationships between Statistics for the images Clicking the Toggle 2D 3D button switches between these representations Show Images Names The data may either be viewed as a set of images spatially placed by similarity a set of names of those images similarly placed or as a sphere Clicking on Show Images Names toggles between these Nearest Neighbours The nearest neighbour to an image is the image that is nearest to that image Nearest may either be in the sense of the 2D or 3D distance in the viewer or in the high dimensional Euclidean distance between statistics vectors Clicking in the Neighbour Join will toggle between showing lines joining nearest neighbours in each of these senses The number of images whose nearest ne
6. ach containing an image class then all images must have entries defining what their class is within that category References 1 Hamilton N Wang J Kerr M C Teasdale R D Statistical and visual differentiation of high throughput subcellular imaging BMC Bioinformatics 2009 to appear 2 Hamilton N Pantelic R Hanson K Karunaratne S Teasdale R D Fast automated cell phenotype image classification BMC Bioinformatics 2007 8 113 3 Hamilton N Teasdale R D Visualizing and clustering high throughput sub cellular localization imaging BMC Bioinformatics 2008 9 81 Reference 1 describes the main methods and testing of iCluster This is the article to reference if you find iCluster useful in your research Reference 2 describes the threshold adjacency statistics that are utilised by iCluster for visualisation and statistical testing Reference 3 describes the method of mapping statistics into 2 or 3 dimensions for visualisation and an early prototype of iCluster iCluster is built using the fantastic proce55ing programming language See http www processing org Problems iCluster runs but is sluggish when viewing image A limiting factor in the visualisation is the graphics card of the computer A poor graphics card or integrated graphics may produce poor results for larger image sets A better graphics card will improve things Alternatively the resolution that images are held in memory at may be reduced by 92
7. ages to by clicking on one of the class checkboxes Deselect All All of the selected images may be deselected by clicking the Deselect All checkbox Selected Images Only If images have been selected via right mouse clicking that set of images may be viewed side by side in 2D by clicking the Selected Images Only checkbox Exit this mode by clicking the checkbox again Image size may be increased decrease in this mode using Right Mouse Button Scroll Mouse Button Load Save Data An image set may be loaded into iCluster by clicking the Load Data button By selecting either a tiff or png format image in a directory using the file browser that appears all of the images in that directory will be loaded into iCluster However images so loaded will all be of the Default class To load images with classes also supplied a Image Description File described below must be in the same directory as the images Selecting this file will load the images their associated classes and any statistics or coordinates previously calculated An image description file may be saved by clicking on Save Data and selecting or creating a file in the same directory as the images The next time iCluster is started the last saved Image Description File will be automatically loaded along with the associated images Statistical Testing Classes of images may be tested against each other to detect if the are statistically different When the
8. hat Java be installed There are 4 directories application where is one of linux windows macosx these contain an executable for each operating system and images which contains an example image set together with an Image Description File icluster_description txt To run iCluster simply enter the appropriate directory for your operating system and start the executable Java version gt 1 5 needs to be installed on the computer Source code is in the subdirectory src To compile the source code you will need the proce55ing software available from http proce55ing org Loading an Image Set To load an image set the images should be placed in a single directory Images need to be 8 bit gray scale in either tiff or png format if statistics are to be generated by iCluster any png image can be displayed in iCluster but statistics generation requires 8 bit gray scale The natural format of iCluster is png and tiff format images are automatically converted to png and saved to disk Images may be assigned to classes for instance treated or untreated in an image description file that is contained within the same directory as the image set Classes are distinguished in the visualiser by border colours of the images visualised The format for the file is given at the end of the document An example image set together with an image description file icluster_description txt can be found in the images folder in the dis
9. iCluster User Manual v1 03beta 2009 Nick Hamilton License iCluster is released under Gnu General Public License It comes with no warranty whatsoever Details of the licence may be found in the licence file contained in the distribution Software 2009 Nick Hamilton n hamilton imb uq edu au Images 2009 Rohan Teasdale r teasdale imb ug edu au If you find this software useful in your research please reference the pare 1 below Overview The core idea of iCluster is to spatially place images of fluorescent protein subcellular localisation is such a way that images that are similar are spatially close It does this by generating threshold adjacency statistics for each image to associate a 27 dimensional vector of real numbers with each image The set of points in 27 dimensional space is then mapped into 2 or 3 dimensions in such a way as to preserve the distances between the points as well as is possible Hence statistically similar images are spatially close in the visualisation Statistical tests for difference between image sets may also be performed statistically representative images found along with a number of other features The methods are described in 1 2 3 Here how to use the principle features of iCluster are described Please email Nick Hamilton n hamilton imb ug edu au with any comments bug reports or requests Installation The distribution comes with executables for Windows Mac OSX and Linux It requires t
10. ighbours are the same class is also displayed This gives a sense of how separated the images classes are Note Results may vary according to whether Normalise Statistics below and Select Statistics above are set Representative Images A representative image for each class may be shown either using the Rep Images Only checkbox where images are shown side by side or Toggle Rep Images where representative images are shown spatially located in 2D or 3D and all other images are hidden A representative image is chosen for each class as that image which is nearest to the centroid of the image statistics for that class Image size may be increased decrease in this mode using Right Mouse Button Scroll Mouse Button Note Results may vary according to whether Normalise Statistics below and Select Statistics below are set Add Class New classes to which images may belong may be added by clicking on the Add Class checkbox When selected the user is prompted to enter a name for the class Press Enter to conclude entering the name One use for this is to create an Outlier class of images which are unusual in some way Reclassify Selected Images that have been selected via right mouse clicking appear with a red tint Such a set may then be reclassified to another class by clicking the Reclassify Selected checkbox Once checked then user can choose which class to reclassify the im
11. irst calculated as follows The normalisation means and standard deviations of the TAS statistics for the 500 images of 10 subcellular localisations in Image Set A described in 1 were calculated For any new image set the TAS as calculated and then each statistic is standardised by subtracting the mean and dividing by the standard deviation for that statistic as calculated for the images for Image Set A The idea is that Image Set A provides a reference by which to standardise the statistics for any image set Hence when standardised the Euclidean distances between images statistics vectors are comparable across any image sets for which the statistics have been standardised In other words standardisation means we have a metric to measure the distance between any images To use normalised rather than standardised statistics select the Normalise Statistics check box Show Hide Image Masks m key Statistics are generated for each image by first selecting the cellular regions in the image using a simple thresholding scheme The threshold is calculated as m 0 90 where m is the modal intensity in the image that is the most common intensity value and o is the standard deviation of the intensities in the image While this produces a reasonable selection for the majority of image iCluster has been tested on it may not work in all cases such as when the background is uneven In such a case the result is often an outlier image Masks may
12. mitochondria which is as before but with 010_mitotracker_mask png as the second entry that being the name of the mask image associated with the image 010 _mitotracker png The fact that the second entry is a mask image is recognized by it s name terminates in either png or tif To supply your own statistics for each image a line might look like 010 _mitotracker png mitochondria 0 12 0 2 1 45 11 2 2 1 7 0 3 7 2 1 4 where the sequence of comma separated numbers are the statistics to use Note Missing values are not supported by iCluster so ensure all lines contain the same number of comma separated real numbers To have multiple categories for the classes of an image see Image Categories above a list of classes may be given instead of a single class For instance 010 _mitotracker png mitochondria not punctate might describe the image 010 mitotrackerpng as being of mitochondria in the subcellular localisation category but also being of class not punctate in the category punctuate not punctate Similarly user supplied image statistics might be appended to such an entry The most important point for the Image Description File is that every line must have exactly the same number of comma separated entries Hence if a mask image is supplied for one image masks must be supplied for all Similarly each image must have the same number of image statistics if supplied and if there are multiple categories e
13. tribution When iCluster starts it will automatically read in the last set of images loaded if an image description file was saved or loaded see below If no image set has previously been loaded when iCluster starts a file selection window will appear Selecting an image description file will then load and display the image set Note If images are missing the will be displayed in iCluster as a black image with a red X Note If an image set is loaded for the first time statistics will be automatically generated for that set if statistics have not been supplied in the image description file However the initial placement of the images in the viewing environment will be random To find an optimal spatial placement of the images so that those that are statistically similar are spatially close click on the Sammon Map check box see below Navigating an Image Set The following mouse and keyboard controls may be used to navigate around an image set View Hide an Image Class Left mouse click on the check box for that class Zoom In Out Mouse scroll Rotate Translate View Right mouse drag Increase Decrease Image Size Right mouse mouse scroll Increase Decrease Sphere Size Left mouse mouse scroll View image in detail Left mouse on image Select an Image Right mouse on image Select a Group of Images Right Mouse on a central image and drag to select around it Show Hide Image Masks m Increase Decrease Image Resolution Snapshot p
14. your own image statistics within iCluster the ability to select de select subsets of them may be useful if you wish to examine the effect on clustering or statistical testing of just using certain of your statistics See Sample Image Description File below for examples of how to use your own statistics in iCluster Image Categories This is probably not of interest to most users but iCluster as well as supporting having a class associated with each image also supports having categories of classes For instance the category Subcellular localisation might contain the classes Nucleus Cytoplasm Mitochondria and so on Most image sets with have just one category containing the classes However it is possible to have other categories of classes and you can switch between these by clicking on the Change Image Categories checkbox Another category might be one that contains the classes Punctate and Not Punctate Hence it is possible to have multiple categories with a class from each category associated with each image and then to switch between views of each category Categories and the classes within those categories may be defined in the Image Description File and an example may be found at the end of this document Sammon Mapping Sammon mapping is the process by which high dimensional statistics vectors are mapped into 2 or 3 dimensions in such a way as to preserve distances between points are well as
Download Pdf Manuals
Related Search