A Graph-Cut Guided ROI Segmentation Algorithm with Lightweight Deep Learning Framework for Cervical Cancer Classification
T L Shiny, Kumar Parasuraman · International Journal of Advanced Computer Science and Applications · 2023
Cervical cancer classification has witnessed numerous advancements through deep learning methods; however, existing approaches often rely on multiple models for segmentation and classification, leading to heightened computational demands and prolonged training times. In this research, a lightweight deep learning framework for cervical cancer classification is presented. The framework comprises three primary components: a Graph-Cut Guided Region of Interest (ROI) segmentation algorithm, a streamlined DenseNet architecture, and a Multi-Class Logistic Regression classifier. The Graph-Cut Guided ROI segmentation algorithm is used to accurately isolate nuclei regions within multicellular Pap smear images. This is a lightweight algorithm that is able to achieve high segmentation accuracy with minimal computational overhead. The streamlined DenseNet architecture is used to efficiently extract salient features from the segmented images. This architecture is specifically designed to reduce feature redundancy and eliminate incongruous feature maps. The Multi-Class Logistic Regression classifier is used to classify the segmented images into different cell types and stages of cervical cancer. Experimental results show the proposed method is able to achieve high classification accuracy with minimal training time. The framework was trained and evaluated on a dataset of 963 Pap smear images. The proposed framework achieved a 98% cell type classification accuracy in precision, recall, and F1-score for classifying multi-cell Pap smear images. The training loss was also very low. The average training time was 21 minutes for different sets of training images, and the average testing time was 0.50 seconds for different sizes of testing images, which is much lower than the existing methods.