Image analysis set-up used at the CJD unit.
The neuropathology of Creutzfeldt-Jakob disease has been classically described as the combination of spongiform change, neuronal loss and astrogliosis. Using immunocytochemical staining PrP plaque formation has also been observed in a minority of cases of CJD. Further immunocytochemical studies have revealed different deposit morphologies in different areas of the brain. While there are similarities between the features of CJD in man and those found in the comparable animal diseases the patterns of pathology found in humans appear to be particularly variable both in targeting and severity. It may be that particular patterns of lesions occur in certain cases which can be related to host genotype or some other factor. The evidence of experimental models of scrapie suggest that a substantial number of strains of the infectious agent influence the pathology observed in cases of this disease.
Before we can determine whether such patterns of pathology exist in CJD it is necessary to assess the natural variability of CJD pathology and also establish a method of relating the pathological features observed in different cases. Towards this goal a substantial investment has been made in the development of image processing techniques for use in the quantification of the different aspects of the pathology of CJD.
We perform analysis on sections of human post-mortem tissue using an image analysis computer system. The analysis is performed using our own software modules specially designed to quantify the different pathological features. The system we use is built around a Sun workstation connected to a modified Leitz microscope. The software we are using has been developed with the assistance of the Medical Research Council's Human Genetics Unit in Edinburgh. The image processing routines are written in the 'C' programming language.
When spongiform change is present in human brain tissue a large number of small holes or vacuoles can be seen in histological samples of the affected tissue, (see figure). The frequency and morphology of these holes are characteristic of this disease and are not seen in normal brain tissue. As yet the cause of this vacuolation and the precise mechanisms involved are unexplained, although, it is thought that the patterns of vacuolation which occur in the brains of individuals affected by this disease may be characteristic of different forms of the disease or different modes of infection. To recognise these different patterns of spongiform change a large number of samples have to be examined as objectively as possible. For this reason we have invested in the use of automated image analysis.
In order to quantify and analyse the patterns of these lesions the computer must be able to recognise these holes as being distinct from other normal tissue spaces in the brain (e.g. capillaries) and artifactual tears and damage caused during tissue processing. The recognition function is performed on digitised microscopic images of conventionally (H & E) stained histological samples.
The procedure used to recognise spongiform change is sequential in nature involving an image intensity thresholding stage followed by object removal, based on morphological constraints, and finally recognition of the actual vacuoles. The output of the system is available in two forms; either an overall value for the spongiform change is supplied as a percentage of the total area of tissue examined or the absolute location of each vacuole is generated as an overlay which can be placed onto an image of the complete tissue sample.
The first stage of the processing algorithm removes all the actual tissue from the image leaving only lighter areas of the image which could represent vacuolation. The thresholding is performed on optical density corrected images, this involves a pre-processing stage which removes any spurious image intensity variations caused by non-linearities in the image capture. On the image scale of 0 (representing transmitted light) to 255 (representing absorbed light) we have used a fixed threshold of 15 intensity units. During the thresholding process any image parts which were darker than this level are removed. The later stages of the algorithm involve constraints on the size and shape of allowable objects. In essence vacuoles are circular in nature and two morphological constraints are used to reject objects which are not fairly circular. The details of the computational stages involved in the rejection process have been published.
Size constraints are also placed on each candidate vacuole to remove tissue spaces which are unlikely to be areas of genuine vacuolation. After any objects which are of the wrong shape and size have been removed then all the remaining objects are measured and summed to produce an overall figure for the amount of spongiform change observed within the present image. Successful spongiform change recognition is shown below.
The actual position of each vacuole is also noted in relation to the tissue sample so that this information can be overlaid to illustrate the distribution of particular areas of vacuolation. The output of this type of analysis on an example section of human cortex is shown below.
In order to examine a particular area of tissue a low magnification image of the tissue section displayed on the computer screen is used to define a region of interest in which we wish to measure the total amount of spongiform change. The software allows the user to define this region of interest as an arbitrary shape and then automatically selects the microscope fields necessary to encompass that entire region. The image analysis system then captures all the individual microscope images necessary to map that entire area by automatically moving the stage to the correct location to capture each field. Following image capture each of the fields are then process individually to locate any spongiform change using the methods described above.
The spongiform change recognition algorithm has been validated by comparison with two neuropathologists performing a subjective rating on an example set of images. In both cases the computer's assessment of the severity of vacuolation was in broad agreement with that of the neuropathologist. We have now applied this technique to the analysis of a considerable number of cases of CJD.
Image analysis was used to quantify PrP plaque formation in the tissue samples of human cerebellum. A counterstain was included in the tissue preparation in order to allow accurate location of the different cerebellar layers. The problem with counterstaining slides for image analysis is that there must be sufficient contrast between the counterstained tissue and the stained deposits for the computer to be able to differentiate between the two quantities. In this case an additional colour filter was used when capturing images in order to enhance the stain/counterstain contrast (blue filter, SP 490nm). A stabilised light source was used to ensure consistency in the data acquisition. In order to remove any other spurious lighting variations background subtraction was performed on all images. This process involves the acquisition of five images representing an area of the present slide outwith the actual tissue section. These images are then averaged together and subtracted from each new image of tissue. This process should correct for any remaining variations in lighting intensity across the imaged field.
In terms of image intensity, the plaque-like deposits present within a sample should appear darker than the surrounding tissue and counterstained cell bodies. The computer requires an intensity threshold in order to be able to differentiate between the dark deposits and the lighter background. It is possible to use automatic methods of setting this threshold in a completely objective way, however, the irregular distribution of the stained deposits visible here meant that a semi-automatic system of intensity thresholding had to be adopted. The setting of the intensity threshold was performed by presenting a very small number of randomly selected microscope fields to a human viewer. Using a visible threshold the human viewer was then asked to vary the value of the threshold until all the stained deposits present within that microscope field were highlighted. The values for the different fields were then averaged to obtain an intensity threshold for that section. In an attempt to overcome any potential operator bias in the thresholding process, all the thresholds used in this study were set by one operator.
With this image analysis system it was possible to use a low magnification view of the tissue sample to manually define an area to be scanned in detail. The stage co-ordinates of the selected fields were stored by the computer and could be recalled at any time during the experiment. Using the image intensity threshold selected earlier the computer then scanned the defined area at a higher magnification and identified all the candidate plaques within that region. As each microscope field was analysed a running total of stained area was updated. Limits were placed on the areas of the largest and smallest plaques in order to avoid the inclusion of artifacts, no further manual intervention was deemed necessary. At the end of the scanning process the computer returned a total value for the overall stained area. The total area of tissue scanned was also available so that percentage area stained could be calculated for each area. The figure shown below gives an example of successful recognition of PrP plaque deposits in human CJD tissue.
This method of quantification of PrP deposition has been applied successfully to plaque like deposits in the cerebellum. These results showed a high correlation between different PrP antisera and also a high level of symmetry between left and right cerebellar hemispheres.
Sutherland K, Rutovitz D, Bell JE, Ironside JW. Evaluation of a novelapplication of image analysis to spogiform change detection. Proceedings of the IEEE International Conference in Imaging Processing, Austin TX, November 1994, pp 378-381.
Sutherland K, Ironside JW. Automatic texure segmentation using morphological filtering on images of the human cerebellum. Proceedings of the IEE International Conference on Image Processing and Its Applications, Edinburgh July 1995, pp777-780.
Sutherland K, Ironside JW. Novel application of image analysis to the detection of spongiform change. Analytical and Quantitative Cytology and Histology 1994; 16(6): 430-434.
Sutherland K, Barrie C and Ironside JW. Automatic quantification of amyloid plaque formation in human spongiform encephalopathy. Neurodegeneration 1994; 3: 293-300.
Sutherland K, Barrie C, Ironside JW. Automatic image analysis of PrP plaque formation in human spongiform encephalopathy. Neuropathology and Applied Neurobiology 1994; 20: 518.
Sutherland K, Macdonald ST, Barrie C, Ironside JW. Assessment of neuropathological targeting in Creutzfeldt-Jakob disease: a quantitative immunocytochemical study. Neuropathology and Applied Neurobiology 1995; 15.
Sutherland K, Goodbrand IA, Bell JE, Ironside JW. Objective quantification of prion protein in spinal cords of cases of Creutzfeldt-Jakob disease. Analytical Cellular Pathology 1996; 10: 25-35.
Sutherland K, Macdonald S, Ironside JW. Quantification and analysis of the neuropathological features of Creutzfeldt-Jakob disease. Journal of Neuroscience Methods 1996; 64: 123-132.
Macdonald S, Sutherland K, Ironside JW. A quantitative and qualitative analysis of prion protein immunohistochemical staining in Creutzfeldt-Jakob disease using four anti prion protein antibodies. Neurodegeneration 1996; 5(1): 87-94.
Macdonald S, Sutherland K, Ironside JW. Prion protein genotype and pathological phenotype studies in sporadic Creutzfeldt-Jakob disease. Neuropathology and Applied Neurobiology (in press) .
This work is supported by the BBSRC.