Multi-Stage CNN with Minimum Duplication Maximum Correlation Method for Pap Smear Images Classification

Abstract

Cervical cancer is a fatal disease that threatens women's lives. It is frequently discovered in an advanced stage and cannot be cured. Cervical cancer is a preventable disease, and one of the most effective methods of prevention is to get regular pap tests. Pap tests detect pre-cancerous cells, which can be treated before they turn into cancer. Generally, a cytopathologist uses a microscope to do Pap smear image analysis. Pap smear images are likely to contain thousands of normal and malignant cells. Due to the cytopathologist's error, the pre-cancerous stages of mild to moderate dysplasia and moderate dysplasia cells are frequently overlooked. In recent times, advanced technology such as automated computer-assisted image analysis (ACIA) has been used to improve the accuracy and efficiency of Pap smear image analysis. The design of ACIA for Pap smear image analysis is mostly based on CNN. In the existing studies based on CNN, emphasis has been placed primarily on accuracy. As a result, increasingly complex CNN architectures with high computational costs for Pap smear image analysis have been developed. The primary goal of this study is to create a CNN model that is both lightweight and computationally efficient for intelligent Pap smear image analysis. The proposed multi-stage CNN architecture utilizes two CNN models. In Pap smear images, the initial CNN model efficiently separates complex cells from the background. The creation of a second CNN model for the extraction of Pap smear image features. This study introduces the Minimum Duplication Maximum Correlation (MDMC) method to reduce redundant and undesired features. This significantly reduces the processing resources that the classification models usually require. Finally, multiclass SVM is used to separate the seven types of normal and abnormal cells from Pap smear images. To achieve a precise accuracy analysis, training efficiency analysis, and computational efficiency analysis of the proposed method, the model is trained and validated using various data set combinations. The experimental results demonstrate that the proposed method is more accurate at classifying seven types of normal and abnormal cells, has very little training loss, and consumes fewer computational resources for classification and model training.