|Year : 2015 | Volume
| Issue : 2 | Page : 97-101
Mathematical (diagnostic) algorithms in the digitization of oral histopathology: The new frontier in histopathological diagnosis
Abhishek Banerjee, Venkatesh V Kamath, RM Lavanya, SK Shruthi, M Deepa
Department of Oral and Maxillofacial Pathology, Dr. Syamala Reddy Dental College, Hospital and Research Centre, Bengaluru, India
|Date of Web Publication||20-Jul-2015|
Venkatesh V Kamath
Department of Oral and Maxillofacial Pathology, Dr. Syamala Reddy Dental College, Hospital and Research Centre, Bengaluru,
Source of Support: None, Conflict of Interest: None
The technological progress in the digitalization of a complete histological glass slide has opened a new door in the tissue based diagnosis. Automated slide diagnosis can be made possible by the use of mathematical algorithms which are formulated by binary codes or values. These algorithms (diagnostic algorithms) include both object based (object features, structures) and pixel based (texture) measures. The intra- and inter-observer errors inherent in the visual diagnosis of a histopathological slide are largely replaced by the use of diagnostic algorithms leading to a standardized and reproducible diagnosis. The present paper reviews the advances in digital histopathology especially related to the use of mathematical algorithms (diagnostic algorithms) in the field of oral histopathology. The literature was reviewed for data relating to the use of algorithms utilized in the construction of computational software with special applications in oral histopathological diagnosis. The data were analyzed, and the types and end targets of the algorithms were tabulated. The advantages, specificities and reproducibility of the software, its shortcomings and its comparison with traditional methods of histopathological diagnosis were evaluated. Algorithms help in automated slide diagnosis by creating software with possible reduced errors and bias with a high degree of specificity, sensitivity, and reproducibility. Akin to the identification of thumbprints and faces, software for histopathological diagnosis will in the near future be an important part of the histopathological diagnosis.
Keywords: Diagnostic algorithms, digitization, mathematical algorithms, oral histopathology
|How to cite this article:|
Banerjee A, Kamath VV, Lavanya R M, Shruthi S K, Deepa M. Mathematical (diagnostic) algorithms in the digitization of oral histopathology: The new frontier in histopathological diagnosis. J Dent Res Rev 2015;2:97-101
|How to cite this URL:|
Banerjee A, Kamath VV, Lavanya R M, Shruthi S K, Deepa M. Mathematical (diagnostic) algorithms in the digitization of oral histopathology: The new frontier in histopathological diagnosis. J Dent Res Rev [serial online] 2015 [cited 2019 Mar 19];2:97-101. Available from: http://www.jdrr.org/text.asp?2015/2/2/97/161216
| Introduction|| |
Analysis of microscopic images of histopathological slides through specific image analysis software has been identified as one of the most potential advancements in histopathology. In traditional systems, pathologists examine biopsies to make diagnostic assessments largely based on cell morphology and tissue distribution. It is always easy to diagnose a case with all classical features, but a lot of variability is noticed while diagnosing cases of dysplasia, certain fibro-osseous lesions or soft tissue tumors. To overcome such problem and improve the histopathological interpretation, it is essential to develop automated tool for the diagnosis.  On the other hand, computational diagnostic tools enable objective judgments by making use of quantitative measures.  The recent advances in pathology have significantly increased the possibility of giving an efficient, standardized diagnosis.
Earlier studies in this field consisted of manual measurements of cell and nuclear morphology; these manual morphometry are often subjected to variation and error.  Automation can improve the practice of pathology by overcoming these limitations of manual microscopy. In automated histopathology, quantitative analysis of histopathological images quantification of features is usually carried out on single cells before categorizing them by classification algorithms such as it is used in scanning software of fingerprint, retina, and iris identification.
Digital pathology has been increasingly encouraged for education, training, clinical practice, as well as in research and diagnosis. The advent of advanced computers, high megapixel based cameras; lenses have allowed pathologists to become untethered from conventional microscopy. The latest image analysis software has allowed the pathologists to render diagnoses with more accurate quantification of histological features. Few worth mentioning advances such as whole slide imaging (WSI), pattern recognition image analysis (PRIA), computer aided diagnosis (CAD) uses a set of algorithms through, which it can create a digitalized replica of the whole slide and then the data can be processed with algorithms to identify features of interest. 
Over many years, a lot of research has been done in regard to histopathological diagnosis as it holds a great promise for the advanced treatment. There are studies which have even showed that automated image analysis solutions have quantified the target protein expression by measuring the immunohistochemical labeling within the various components of the cell and formulation of specific algorithms has helped in attaining this level. There are various advantages by making histopathology automated such as firstly, speeding up the diagnosis with the help of computers (CAD) by pathology detection and grading it, second, it facilitates interactive session which involves the large sized virtual e-slides and lastly it increases the level of accuracy in diagnosis when the automatically generated findings from the slides are combined with the expertise of the histopathologist.  Automation in slide diagnosis aims at the 100% accuracy and lesser false alarm rate. The digitization of slide diagnosis involves the glass slide to get scanned under a slide scanner and it acquires a digital presentation of the complete glass slide including the specific area of display in the slide at any magnification within the limitations of light optics. ,,
This paper will mainly focus on the various steps and algorithms involved in generating an automated diagnosis from the glass slide and the various methods that aims in facilitating and analyzing the slides by automatically detecting the relevant and irrelevant areas with the help of mathematical algorithms and deliver a speedy and accurate diagnosis.
| The Methodology of Delivering Automated Diagnosis|| |
To reach a definitive automated diagnosis, series of logarithmic steps need to be followed as discussed in [Figure 1]. The glass histology slide is obtained which is converted into an image by the means of an automatic slide scanner. The automatic slide scanner works on the principle of WSI. The main aim behind scanning the whole slide is to determine the level of malignancy or dysplasia by examining the slide at a mega scale. The first step in automation is feature extraction (cellular level and tissue level). The automated diagnosis of slides includes the three major steps that are preprocessing, feature extraction and diagnosis. Based on these three computational steps diagnosis is reached which targets at distinguishing the type of the lesion and simultaneously grading the lesion based on the altered features.
|Figure 1: The basic steps in automated diagnosis of histopathological digitized images|
Click here to view
The main target of this step is to identify few specific focal spots or areas in the image.  It is done by reducing the noise from the image prior to the focal area spotting. Noise arises from the uneven distribution of the stain and from processing errors. The most common way of reducing the noise is by thresholding the pixels of the image. Filters especially reduces the noise and improves the thresholding.  Another method of noise reduction is by using mathematical morphology, which is a set theoretic approach that considers pixels in an image as the elements of a set. ,, Once the noise removal is done, the intensity distribution is looked in the smoothened image after this steps the light grained structures are segregated from the dark ones. The area values are calculated for the segregated ones and then it is subtracted from the main area. There are two basic transformations that are also done, that is., opening and closing. Opening increases the sharpness of the image by eliminating the small objects while closing is dilating the image by filling up small gaps. 
The cell segmentation is basically done in order to separate out the cells from the background that is the stroma and thus helps to study them in particular. This is done by two methods, the first method is pixel based and it recognizes whether the pixels from the cell or from the surrounding background stroma and thus identifying the cell while the other method uses the boundary delineation by manually marking the boundary of the cell [Figure 2]. In case of the overlapping cells, specific algorithms like the mathematical morphology , and the watershed algorithms come into play. , Watershed algorithms are useful in detecting the boundary lines between the touching cells. This algorithm splits the image area into disjoint regions, and the watershed lines are closed curves. Watershed algorithms are very sensitive to noisy pictures, and hence it becomes very difficult to select relevant areas, so smoothening of the image becomes very essential.  The median filters preserves the borders of the segments, which is beneficial for the proper feature extraction.  The limitation of these algorithms is mainly over-segmentation, but the use of low-pass filters prior to segmentation lessens the over-segmentation problem. The k-means learning algorithms automatically clusters the pixels, according to their color information. 
The extraction of the features, which are frequently known as deviations of the sample are done at cellular and tissue levels. The features are mainly extracted by quantifying the changes which are observed. To measure the features, various parameters are considered like the morphological, textural, fractal, and intensity within the cellular and tissue boundaries.  The morphological features provide information about the size and shape of the cell. There are parameters, which defines the morphology such as radius, area, perimeter, compactness, axis, symmetry, and concavity. , The fractal based features give information on the regularity and complexity of an object by quantification its self-similarity level. Textural features provide information about the variations in the intensity of the surface by characterization of smoothness and coarseness. The Voronoi diagrams and the Delaunay triangulations help in defining the area and shape of the cells in their polygons. These algorithms work by a principle of the interconnecting network that is joined together and helps to identify the topological features. ,
The diagnosis is followed after the appropriate noise reduction, and cellular segmentation are done, features are extracted. In this particular step, statistics plays a very important role in determining whether the features, which are selected for diagnosis holds any significant difference from the other features or not. There are machine learning algorithms which help in differentiating these specific features; the relevant from the nonrelevant ones. The most commonly used are k-nearest neighborhood algorithms, logistic regression method, fuzzy systems, linear discriminate functions, and decision trees. 
| Pattern Recognition|| |
To recognize the pattern from the WSI, the image is first tessellated. The tessellation lines divide the WSI into numerous blocks. The size of the blocks depends upon the applied pattern recognizing algorithm. Each block is added with one additional m-pixel, which also serves as a boundary, and it helps in overlapping with the additional block so that while reading the stain continuity is maintained, and there are no sharp boundaries between the stain [Figure 3].  The center of the each staining region should be counted in order to avoid double counting. There are automated PRIA software which have been evaluated in terms of microscopic assessment through WSI. A study was carried out on microscopic assessment of metastatic lung cancer comparing the automated pattern identifying image analyzer and manual morphometric image segmentation. PRIA was found to be better for tissues with limited phenotypic diversity.  There were few limitations encountered while working with the algorithm are image file size and image processing errors.
|Figure 3: Whole slide pattern recognition by tessellation. The slide image is divided into "n × n" sized blocks to enable recognition and after that each block is patched with additional "m" width margin|
Click here to view
| Tissue Based Diagnosis Related Algorithms|| |
The derived algorithms which are to be applied for automated diagnosis include both textural and objective information acquisition techniques. The image undergoes a series of texture analysis through various transformations such as thinning, Hough, Prewitt, and Fourier operational methods.  External information is required for the segmentation process to identify the objects which are searched for. Most of the objects can be segmented which are within the field of view using dynamic thresholding and external knowledge. The spatial relationships between the objects can be analyzed by syntactic structure analysis, which is based on the neighborhood algorithm.  The whole image of the slide which consists of numerous pixels is very difficult to assess at a particular time and hence the image is stratified for analysis. Nyquist's formula can be applied to calculate the minimum percentage of analyzed image areas from the whole.  The analysis of the complete slide virtually also ensures an accurate diagnosis by detecting relevant areas using these algorithms [Figure 4]. 
| Stain Recognition Algorithms|| |
Stain recognition algorithm is composed of three important steps that perform pixel-wise color classification, morphological smoothening and analysis of isolated stained region. The red green blue color channels are used for the classification features and the recognized output were usually classified as either 0 (background) or 1 (foreground) and the stain classification is done by using nearest neighborhood method. , This algorithm classifies each image pixel based on its color differences with the pixels in the trained images. The minimum difference will determine whether it is foreground stain or background stain. The classification is applied to the input image, and a binary output is obtained. There might be small fragments in the output, which are generally classified as errors, but are easily removed by the morphological opening and closing mechanisms. After the strain localization is done, the image is subjected to density estimator. 
Since the localization of the stains are scattered over the whole slide, kernel density estimator along with the regression analysis software is used to calculate the density distribution.  Finally, after the density estimation is done and features are selected, the image is subjected for validation. The stain recognition algorithm is shown in [Figure 5]. 
| The Various Applications of Using Algorithms for Automated Slide Analysis|| |
The use of mathematical based algorithms has given histopathology based diagnosis a new edge. There are various applications of automated slide analysis such as: 
- Telepathology - there are three major ways by which telepathology is modulated like storage and forwarding of images, robotic dynamic microscopy, and virtual slides. The digitization plays a very important role in telepathology. The virtual slides are sent through the high-speed internet and helps in histopathology networking. This also helps in reducing consultation time and bias. There are few limitations like the inappropriate field selection, lack of depth perception, time lag, set up costs etc., but it helps in the better allocation of pathologists in remote areas 
- WSI - this includes the scanning and image acquisition, image quality and focus, three-dimensional view, slide navigation and graphical user interface. WSI has technologically revolutionized the traditional routine microscopy. In this process, the entire pathological glass slide is converted into an image file, which can be easily shared by pathologist across the globe. The whole digital slide can be viewed under different magnifications. WSI also helps in slide archiving, continuing medical education, and consultation 
- Digital image analysis - It helps to acquire data with higher resolution. It also plays a very crucial role in the fractal analysis of scaling relations. The main disadvantage is the costing of the application and the inconvenience which is caused due to the huge data collection
- CAD and enhancing pathology consultation networks.
The above facts about digitization make it very clear that pathology informatics will continue to grow. These algorithms and software are useful, and it plays a pivotal role in rendering diagnosis along with the traditional microscopic systems. Digitization will not only render a diagnosis but from an optimistic view it will also actively take part in treatment planning in the near future.
| Conclusion|| |
Digital pathology is a boon of scientific innovations which has transformed the practice of pathology. The advances in microscopy and computer reflect the qualitative improvement of histopathological image acquisition. There is a trend of steep increase in the number of patents accepted in the field of virtual microscopy. A lot of efforts have been made to accurately render the diagnosis, as well as to classify the lesion based on the progression, but there are even too many challenges to evaluate the reliability of the designed diagnostic systems. The improper use of the algorithms and evaluation methods can even lead to misleading results and often leads to bias. The innovation of these algorithms has made it easy to identify the diagnostic relevant area from the irrelevant ones with a higher degree of accuracy. It has also served as a pruning step in computer assisted diagnosis in setups, and, as well as in the telepathology workstations. Algorithms should be cross-validated prior to customization and application, and so it can deliver greatest benefits to the pathology consultation. Hence, it can entail an ongoing collaboration between the computer technologists and the pathologists to deliver a better and accurate diagnosis so that the treatment can be better planned with much more ease.
| References|| |
Andrion A, Magnani C, Betta PG, Donna A, Mollo F, Scelsi M, et al.
Malignant mesothelioma of the pleura: Interobserver variability. J Clin Pathol 1995;48:856-60.
Demir C, Yener B. Automated Cancer Diagnosis Based on Histopathological Images: A Systematic Survey. Technical Report TR-05-09, Computer Science Department at Rensselaer Polytechnic Institute; 2005.
Cucoranu IC, Parwani AV, Vepa S, Weinstein RS, Pantanowitz L. Digital pathology: A systematic evaluation of the patent landscape. J Pathol Inform 2014;5:16.
Bahlmann C, Patel A, Johnson J, Ni J, Chekkoury A, Khurd P, et al
. Auto-mated detection of diagnostically relevant regions in H&E stained digital pathology slides. Proc SPIE Med Imag 2012;83:504-09.
Leong FJ, McGee JO. An automated diagnostic system for tubular carcinoma of the breast - An overview of approach and considerations. Stud Health Technol Inform 2003;97:57-72.
Molnar B, Berczi L, Diczhazy C, Tagscherer A, Varga SV, Szende B, et al.
Digital slide and virtual microscopy based routine and telepathology evaluation of routine gastrointestinal biopsy specimens. J Clin Pathol 2003;56:433-8.
Weinstein RS. Innovations in medical imaging and virtual microscopy. Hum Pathol 2005;36:317-9.
Vandana BS, Antony PJ. Automated segmentation using histopathology images as a diagnostic confirmatory tool in detection of bone cancer. Int J Comput Appl 2012;57:48-55.
Otsu N. A threshold section method from gray level histograms. IEEE Trans Syst Man Cybern 1979;9:62-6.
Serra J. Image analysis and mathematical morphology. Cytometry 1983;4:184-5.
Malpica N, de Solórzano CO, Vaquero JJ, Santos A, Vallcorba I, García-Sagredo JM, et al.
Applying watershed algorithms to the segmentation of clustered nuclei. Cytometry 1997;28:289-97.
Vincent L, Soille P. Watersheds in digital spaces: An efficient algorithm based on immersion simulations. IEEE T Pattern Anal 1991;3:583-598.
Gunduz C, Yener B, Gultekin SH. The cell graphs of cancer. Bioinformatics 2004;20 Suppl 1:i145-51.
Street WN, Wolberg WH, Mangasarian OL. Nuclear feature extraction of breast tumour diagnosis. Int Symp Electron Imaging 1993;1905:861-70.
Spyridonos P, Ravazoula P, Cavouras D, Berberidis K, Nikiforidis K. Computer based grading of H and E stained tissue sections of urinary bladder carcinomas. Med Inform Internet Med 2001;26:179-90.
Kayser K, Radziszowski D, Bzdyl P, Sommer R, Kayser G. Digitized pathology: Theory and experiences in automated tissue-based virtual diagnosis. Rom J Morphol Embryol 2006;47:21-8.
van Diest PJ, Kayser K, Meijer GA, Baak JP. Syntactic structure analysis. Pathologica 1995;87:255-62.
Yeh FC, Ye Q, Hitchens TK, Wu YL, Parwani AV, Ho C. Mapping stain distribution in pathology slides using whole slide imaging. J Pathol Inform 2014;5:1.
Webster JD, Michalowski AM, Dwyer JE, Corps KN, Wei BR, Juopperi T, et al
. Investigation into diagnostic agreement using automated computer-assisted histopathology pattern recognition image analysis. J Pathol Inform 2012;3:18.
Kayser G, Riede U, Werner M. Towards an automated morphological classification of histological images of common lung carcinomas. Electron J Pathol Histol 2002;8:22-30.
KAYSER K, KAYSER G, Virtual microscopy and automated diagnosis. In: GU J, Ogilvie R, editors. Virtual microscopy and virtual slides in teac hing, diagnosis and research Taylor and Francis, Boca Raton, 2005.
Wu X, Kumar V, Quinlan JR, Ghosh J, Yang Q, Motoda H, et al
. Top 10 algorithms in data mining. Knowl Inf Syst 2008;14:1-37.
Parzen E. On estimation of a probability density function and mode. Ann Math Stat 1962;33:1065-70.
Gabril MY, Yousef GM. Informatics for practicing anatomical pathologists: Marking a new era in pathology practice. Mod Pathol 2010;23:349-58.
Wibur DC, Kalil M, Colvin RB, Duncan LM, William CF, Ferry JA, et al
. Whole slide imaging digital pathology as a platform for teleconsultation: A pilot study using paired subspecialist correlations. Arch Pathol Lab Med 2009;133:1949-53.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5]