Diagnostic Accuracy of Entropy Based Image Analysis of Cervical Precancerous Cells.

The objective of the study was to demonstrate the modified Bragg-Williams type analysis of digital microscopic images is an analytically valid methodology for classifying squamous cervical cytology. We analyzed 5318 single cell cervical cytology images from the SIPaKMeD and Herlev databases to calculate a Bragg-Williams order parameter (S2) using a pixel intensity histogram-based approach. A classification model for S2 was created to differentiate normal, LSIL and HSIL. A set of test images from different databases were blindly classified using defined ranges to determine the sensitivity, specificity, and areas under the curve. Results showed distinct ranges of S2 for Normal: 0.7181-0.8633 SD 0.1452, LSIL: 0.580-0.6626 SD 0.0826, and HSIL: 0.3289-0.550 SD 0.2211, with dysplastic cells exhibiting lower degrees of order. The S2 classification model was found to have 100% sensitivity and specificity over the sample set, demonstrating the suitability of the methodology for cervical cytology classification.