Machine Learning Based Approach for Automated Cervical Dysplasia Detection Using Multi-Resolution Transform Domain Features

Abstract

Pattern detection and classification of cervical cell dysplasia can assist with diagnosis and treatment. This study aims to develop a computational model for real-world applications for cervical dysplasia that has the highest degree of accuracy and the lowest computation time. Initially, an ML framework is created, which has been trained and evaluated to classify dysplasia. Three different color models, three multi-resolution transform-based techniques for feature extraction (each with different filters), two feature representation schemes, and two well-known classification approaches are developed in conjunction to determine the optimal combination of “transform (filter) ⇒ color model ⇒ feature representation ⇒ classifier”. Extensive evaluations of two datasets, one is indigenous (own generated database) and the other is publicly available, demonstrated that the Non-subsampled Contourlet Transform (NSCT) feature-based classification performs well, it reveals that the combination “NSCT (pyrexc,pkva), YCbCr, MLP” gives most satisfactory framework with a classification accuracy of 98.02% (average) using the F1 feature set. Compared to two other approaches, our proposed model yields the most satisfying results, with an accuracy in the range of 98.00–99.50%.