Contextual analysis and intermediate cell markers enhance high‐resolution cell image analysis for automated cervical smear diagnosis

Abstract

Until now, efforts to automate cervical smear diagnosis have focused on analyzing features of individual cells. In a complex specimen such as that obtained from a cervical scrape, diagnostically significant cells may not be adequately represented or may elude detection by the automated technology. An approach is needed that extracts additional quantitative information from cervical smears beyond what the cell-by-cell approach can provide .A new methodology, contextual analysis, was developed to extract global quantitative information about cells, cell clusters, and background debris. This pilot study was designed to compare the efficacy of contextual analysis with highresolution, single cell analysis and the analysis of intermediate cell markers. Thirty-four samples prepared as monolayers and stained with the FeulgenThioninKongo Red stain were measured. Contextual analysis alone was able to classify 91% of the smears correctly: single cell analysis classified 94% of the cells correctly; and the intermediate cell analysis correctly identified the smear diagnosis for 84% of the cells. When all three analysis methods were combined into a simple smear level classifier, the overall smear classification accuracy was improved over those obtained using the three methodologies alone.

Exploratory analysis of quantitative histopathology of cervical intraepithelial neoplasia: Objectivity, reproducibility, malignancy‐associated changes, and human papillomavirus

“…To our knowledge this is the first result on MACs in cervical histopathology, although MACs have been previously observed in the cervical cytology literature (19 -21). These results agree with findings reported in the literature on cervical biomarkers (20) and the literature on MACs in other organ sites (22,23): lung (24 -26), colon (17), stomach (27), and oral cavity (28).…”

“…The discriminant functions were constrained to use no more than three features by selection of the F to enter and the F to remove. (20) Low, medium, and high DNA amount Low, medium, and high DNA area Low, medium, high and medium_high DNA compactness Low, medium, high and medium_high DNA average distance Low, medium, and high density object Low, medium, and high centre mass Low_vs_medium, low_vs_high and low_vs_medium-high DNA Markovian texture (7) Entropy, energy, contrast, correlation, homogeneity, cl_shade, cl_prommence Non-Markovian texture (5) Density_light_spots, density_dark_spots, center_of_gravity, range_extreme, range_average, Fractal texture (3) Fractal_area1, fractal_area2, fractal_dimension Run length texture (20) Short_runs_mean, short_run_stdv, short_run_min, short_run_max long_runs_mean, long_run_stdv, long_run_min, long_run_max gray_level_mean, gray_level_stdv, gray_level_min, gray_level_max run_length_mean, run_length_stdv, run_length_min, run_length_max run_percent_mean, run_percent_stdv, run_percent_min, run_percent_max (18) Area (mean), area (standard deviation), area (skewness), area (kurtosis), area disorder, perimeter (mean), perimeter (standard deviation), perimeter (skewness), perimeter (kurtosis), roundness factor (mean), roundness factor (standard deviation), roundness factor (skewness), roundness factor (kurtosis), roundness factor heterogeneity, number of sides (mean), number of sides (standard deviation), number of sides (skewness), number of sides (kurtosis) Features derived from Delaunay graph (4) Nearest-neighbor distance (mean), nearest-neighbor distance (standard deviation), Delaunay nearestneighbor distance (mean), Delaunay nearest-neighbor distance (standard deviation) Features derived from the minimum spanning tree (MST) (7)…”

Detection of Malignancy Associated Changes in Cervical Cell Nuclei Using Feed‐Forward Neural Networks

“…It has been observed that there are significant changes in the quantitative nuclear features of intermediate cells from cervical smears of different grades [2,4,11,19]. These nuclear changes have been used to design classifiers using linear discriminant functions (LDF) [12,14,15,20]. Nonlinear classifiers such as neural networks have already been successfully applied to identifying diagnostic cells in cervical smears [3,17].…”

A Review of Caveats in Statistical Nuclear Image Analysis

“…Without testing, the probability will be high that the results confuse rather than inform. A number of studies have been published on nuclear image analysis in diagnosis and prognosis of cancer without a well-defined test set [1,27,28,[36][37][38]40,41,50,51,65,69,75,83,89,95,99,112,115,122,127,134,137].…”

“…Therefore, image analysis methods have been developed in an attempt to obtain more objective diagnosis. Diagnostic systems using image analysis have been applied to many different organs, including the uterine cervix [41,49,80,90,102,131,133,137], the ovaries [30,66], the breasts [9,26,115,119], the liver [37,53,67], the thyroid gland [18,31,68,76,110], the lungs [71], the kidneys [124], the brain [27], the pancreas [105], the colon [89], the oral cavity [114], the skin [38,74] and lymphoid tissue [91]. Nuclear image analysis methods are still in an investigative phase, and although promising achievements have been made, their importance for diagnosis in clinical use has still to be shown.…”