- Top of page
- MATERIALS AND METHODS
- LITERATURE CITED
- Supporting Information
Automated and quantitative histological analysis can improve diagnostic efficacy in colon sections. Our objective was to develop a parameter set for automated classification of aspecific colitis, ulcerative colitis, and Crohn's disease using digital slides, tissue cytometric parameters, and virtual microscopy. Routinely processed hematoxylin-and-eosin-stained histological sections from specimens that showed normal mucosa (24 cases), aspecific colitis (11 cases), ulcerative colitis (25 cases), and Crohn's disease (9 cases) diagnosed by conventional optical microscopy were scanned and digitized in high resolution (0.24 μm/pixel). Thirty-eight cytometric parameters based on morphometry were determined on cells, glands, and superficial epithelium. Fourteen tissue cytometric parameters based on ratios of tissue compartments were counted as well. Leave-one-out discriminant analysis was used for classification of the samples groups. Cellular morphometric features showed no significant differences in these benign colon alterations. However, gland related morphological differences (Gland Shape) for normal mucosa, ulcerative colitis, and aspecific colitis were found (P < 0.01). Eight of the 14 tissue cytometric related parameters showed significant differences (P < 0.01). The most discriminatory parameters were the ratio of cell number in glands and in the whole slide, biopsy/gland surface ratio. These differences resulted in 88% overall accuracy in the classification. Crohn's disease could be discriminated only in 56%. Automated virtual microscopy can be used to classify colon mucosa as normal, ulcerative colitis, and aspecific colitis with reasonable accuracy. Further developments of dedicated parameters are necessary to identify Crohn's disease on digital slides. © 2008 International Society for Analytical Cytology
The urgent need for the increase of histological diagnostic efficiency requires the support of automated, computerized prescreening systems.
The ability of computers to render accurate diagnoses on cytopathologic specimens such as cervical Papanicolaou smears is well established and well documented since the early 1980s by Wittekind et al. (1) and Stenkvist et al. (2).
For many reasons, automated analysis of histologic sections is profoundly more difficult than the automated analysis of cytopathologic preparations. Histologic sections may have complex architecture or high cellularity, whereas cytologic preparations have relatively simple architecture and relatively low cellularity. Furthermore, histologic sections are prone to artifacts such as chatter, folding, contamination, fragmentation, thermal injury, and crush-related injury. These artifacts represent noise that automated analysis must ignore during the final interpretation.
For automated analysis to have maximal clinical utility, higher order analytical functions such as precise architectural measurements of glands, epithelial surface is needed. Furthermore classification of cells into the proper morphological type, evaluation of the fine nuclear and cytoplasmic detail, and even different types of stroma must be performed.
Despite these obstacles, recent studies demonstrate highly effective automated analysis of histological sections, including the detection of cancer cells (3). Most studies have focused on routinely processed hematoxylin-and-eosin-stained sections. The use of automated analysis has been successfully extrapolated to quantitative morphometry of immunostained sections in the setting of mammary carcinoma. Francis et al. (4) introduced an analysis method for estimation of PCNA in breast carcinoma which worked on single field of views with high accuracy.
Thompson et al. (5) presented the knowledge-guided segmentation method that partitioned colorectal images to different histologic components where glands were recognized with 85% accuracy. The knowledge-guided method was adopted for prostate samples by Bartels et al. (6) and was shown that measurement of progression or regression is possible by detecting and analyzing prostatic lesions. Hamilton et al. (7) introduced an image texture analysis method to locate dysplastic fields in colorectal samples. The automatic identification of focal areas of colorectal dysplasia was based on co-occurrence matrix and optical density at low power microscopic images. This study also showed that the combination of automated localization at low magnification and knowledge-based image segmentation at high magnification creates an automated tool for supporting diagnostic decision making. Esgiar et al. (8) presented a new solution for colon carcinoma identification based on geometric and texture analysis and achieved 90% accuracy in the classification. Four years later Esgiar et al. (9) extended their measurements with fractal analysis which increased the accuracy to 95%.
All of the above-mentioned studies showed good results but they have the same limitation in that these methods use single images. The analysis of image information available on the whole slide can supply essential additional data. For this reason, recent studies present automated histological analysis based on whole slide imaging. Petushi et al. (10) introduced new grade-differentiating parameters for breast cancer by examining the whole digital slide, providing an opportunity for pathologists to support their diagnosis by objective quantitative measurements.
Clearly, more sophisticated algorithms are required for automated analysis to address the issue of automated disease classification. Furthermore, these algorithms must incorporate the analysis of large structures on large field of views on an entire slide. For example, quantitative analysis of cells throughout an entire slide or for a specific region on a slide that harbors glands might help characterize different diseases. The recent advances in whole slide imaging and visualization support the development of dedicated organ specific algorithms that can be selectively applied on the entire specimen in low or high resolution depending on the analysis request. Therefore, our aim was to develop efficient algorithms to detect and measure higher order architecture and nuclear and cellular alterations in parallel and determine whether these algorithms could be used by automated analysis to reliably diagnose colon mucosa as normal or diseased on digital slides in the environment of virtual microscopy. As attempts toward automated analysis of malignant alterations were performed, our aim was the detection of IBD and the differentiation of IBD in colon form aspecific colitis and healthy mucosa.
- Top of page
- MATERIALS AND METHODS
- LITERATURE CITED
- Supporting Information
The quantification of histologic parameters for automated quantification and classification is a longstanding problem. Early attempts were made on selected microscopic field of views. Depending on the selected problem, low or high resolution images could only be collected. A comprehensive analysis where architectural features could be seen at low and intermediate magnification (up to 200×) and cytologic features could be seen at high magnification (400×) could not be performed. Because of these limitations and working on single images, previous studies of image analysis concentrated on selected and specific problems of image analysis such as nuclear detection (15), detection of immunopositivity (16), quantification of vascularization (3), mitotic counting (11) in high resolution.
Low resolution analysis of histologic architecture was attempted by Thompson et al. (5) and Esgiar et al. (8) for colonic carcinoma by different groups. An interesting approach of image analysis was introduced by Hamilton et al. (7) for finding dysplastic fields in colorectal sections, using neural networks on a mosaic of pixilated images without any image analysis or image segmentation.
Digital slides offer a flexible platform for image analysis. Using this digital media the same specimen can be morphometrically quantified for architectural components in the 50–200 μm range, similar to the glands, surface, follicles, and for cytologic components (nuclei, cytoplasm) in the 5–10 μm range. Using this media the development of tissue-specific parameters can be started. In our opinion morphometric and densitometric parameters of the sections must be comparable to each other in a standardized environment where the thickness of sectioning, staining, and coverslipping is automated and standardized.
We attempted to develop overall tissue-specific parameters and define the frontiers where additional alteration-related parameters are required. Colon biopsy specimens are ideal for this purpose because they can be frequently obtained and thus provide a major load for the practice. As this study is a preliminary one, we did not include all of the available diagnostic groups. As previous studies showed the feasibility of detection of malignant alterations in colon, we concentrated on a new field of the inflammatory bowel diseases. Three groups were involved in this study namely the aspecific colitis, the ulcerative colitis, and the Crohn's disease.
The correctness of nucleus segmentation is difficult to calculate for the whole project because it would require counting about 2.5 million nuclei. Only limited regions were verified in our study.
The epithelium detection was worse than in our previous gastric study (17) because the epithelial cell characteristic is less intensive and the cytoplasm thickness is smaller.
The cell morphometric features showed no significant differences between the groups. This is understandable as our samples were all benign, containing healthy normal cells.
By discriminant analysis the number of measured parameters could be decreased dramatically from 52 to 8 which is adequate to define a reliable discriminant function for colon classification. Two measured features should be emphasized, gland shape factor and the ratio of cell number in glands and in the whole slide. Both of these parameters have been already utilized by pathologist but they were not automatically quantified until now. The study showed that tissue cytometric parameters can distinguish major disease groups in colon samples.
This study is a good continuation of previous studies on automatic diagnosis. Esgiar et al. (9) reached 95% of colon carcinoma identification and the authors' previous study (17) on classification of gastric biopsies achieved 100% accuracy in the classification of adenocarcinoma and had 86% overall accuracy. The current study tried to concentrate on other than the above-mentioned diseases and performed 88% overall classification rate where the normal from the inflammatory samples were identified 96% correct.
It is also clear that Crohn's disease requires further research to be able to eliminate it from ulcerative colitis. For these type of analyses further intraepithelial, intraglandular analytical algorithms will be needed. In addition, specific granuloma detection algorithm must be developed.
The analysis algorithms could run after scanning automatically on the digital slides. The data and the graphical overlay and masks of the analysis results could be evaluated in the virtual microscopes even in a distance.
Digital slides and virtual microscopy opens new dimensions in histopathology which allows us to analyze whole specimens rather than only a single field of view. The newly developed tissue cytometry features (as a product of virtual microscopy) are new milestones in histo/cytometry and are essential in the classification of colon samples. The study shows that there are strong differences among major colon disease groups, and that a high percentage of such cases can be correctly classified based on tissue and cellular parameters.