A Novel Cervical Cancer Detection Using Contrast Overlap Feature-Based Segmentation with Neural Network Classifier

Abstract Cervical cancer is another prime cause of women’s casualties due to its detection and diagnosis in advanced stages. Modern computer-aided processing exploits magnetic resonance imaging, computed tomography, positron emission tomography, and advanced artificial intelligence to improve early detection of this cancer. In this study, a Contrast Overlap Segmentation Method (COSM) was introduced and described. The extracted features were first classified as high/low based on their contrast intensity to classify different regions. The classification was further extended using a neural network with two hidden layers for accuracy improvement and training initialization. The neural network outcomes were normalized using the ReLU function to identify overlapping and non-overlapping regions to ensure that the infected region in any intensity region was identified. The neural network was trained using a labeled dataset to ensure that high-region segmentation was performed with better sensitivity. Thus, the training process was augmented for segmented and input training sets with better pixel feature classifications. In results section, experimental analysis is using CCAgT cervical cancer dataset. This dataset provides 9339 smear images of cervical cancer cells under 15 slides. For the highest classification rate, the proposed COSM improved the accuracy, precision, and sensitivity by 10.45, 9.62, and 12.38%, respectively. The mean error and computing time were reduced by 9.77 and 7.85%, respectively. The proposed method experienced overhead when handling variable feature distributions. Such distributions require a maximum-to-minimum grouping-based normalization, which is lacking in this method.