Multi-class nucleus detection and classification using deep convolutional neural network with enhanced high dimensional dissimilarity translation model on cervical cells

Advanced cervical screening via liquid-based cytology (LBC)/Pap smear is a highly efficient precancerous cell detection tool based on cell image analysis, in which cells are classified as normal/abnormal. This paper outlines the drawbacks by introducing a new framework for the accurate classification of cervical cells. The proposed methodology comprises three phases: segmentation, localization of nucleus, and classification. In the segmentation phase, we develop a hybrid system that incorporates two binary image patches obtained by a 19-layered convolutional neural network (ConvNet) model with an enhanced deep high dimensional dissimilarity translation (HDDT) based conspicuous segmentation. To get the relevant information from binary patched images, a technique called optimum semantic similarity selective search (OSS-SS) is proposed that returns the localized RGB patched image. A pre-trained ResNet-50 model is retrained using transfer learning on localized patched images in the classification phase. Following that, the selected features from the average pool and fully connected layers are down-sampled using the t-distribution stochastic neighbor embedding (t-SNE) approach. Finally, these combined features are fed into a multi-class weighted kernel extreme learning machine (WKELM) classifier via a sparse multicanonical correlation (SMCCA) method. Three datasets (SIPaKMed, CRIC, and Harlev) are used to evaluate the segmentation and classification task. The proposed approach obtained an accuracy of 99.12 %, specificity of 99.45 %, sensitivity of 99.25 % with an execution time 99.6248 on SIPaKMed. The experimental analysis indicate that our model is more effective than existing techniques.