Cervical Cancer Detection using Deep Neural Networks and Ensemble of Decision Trees

Abstract

Automated cervical cancer screening is an efficient cell imaging based cancer detection application of pattern classification that uses liquid-based cytology (LBC) and Pap smear images. LBC and Pap smear images contain cells which can be categorized into “normal” and “abnormal” categories. Screening system uses segmentation approaches for feature extraction for successful classification. However, successful classification depends on how accurately segmentation is done. In this thesis, an auto-assisted cervical cancer screening without prior segmentation of cervical cells is proposed. Transfer learning approach is used for fine tuning of the new Convolutional Neural Network (CNN), i.e. weights of the convolutional and pooling layers of a pre-trained CNN are transferred to new CNN. Fully connected layers of the new CNN are initialed with values from gaussian distribution. New CNN is then fine-tuned on the cervical cell dataset to learn new weights. Performance of the CNN-based screening system is tested on Herlev dataset for two class problem and seven class problem. Herlev cervical cell dataset consist of seven class data, while two class problem is achieved by combining three normal classes i.e. superficial, intermediate and columnar epithelial as normal class and four abnormal classes i.e. mild dysplasia, moderate dysplasia, severe dysplasia and carcinoma as one abnormal class. A distinguished feature of the proposed approach is that, it achieves its objective without getting into conventional segmentation approach for feature extraction. The immediate impact of this approach can be observed on the classification accuracy of the system. Three different classification approaches are used for comparison analysis on the classification accuracies i.e. softmax, SVM and tree ensemble. Classification accuracies of softmax, SVM and tree ensemble for two class problem is 98.8%, 99.10% and 99.23% respectively. For seven class problem, classification accuracies of softmax, SVM and tree ensemble are 97.21%, 98.12% and 98.85% respectively. These results shows that the proposed system yield better performance in all metrics i.e. accuracy sensitivity and specificity than its previous counterparts as the previous best classification accuracies are 98.3% for two class problem and 96.6 for seven class problem.