Tagging of Uterine Cervix Cases at Cell and Slide Level Through Transfer Learning

In this paper, two attempts have been made to classify the cells of uterine cervix cancer at the cell level and the slide level. The cell-level classification has two-class and a detailed seven-class classification. In case of slide-level classification, an individual slide is assigned with the label of normal/abnormal. In traditional machine learning algorithms, feature engineering was the biggest hindrance to the accurate representation of the cells. Deep Neural Networks consisting of many layers in the middle can extract thousands of feature values from the ground truth, which becomes sufficient for efficient classification. In this study, Deep Neural Networks are trained using the concept of transfer learning, wherein only the parameters of the last layer of the network are tuned. Transfer learning refers to an approach in deep learning where a model trained on one task is reused as the starting point for a model on a second related task. Transfer learning is a very less time-consuming approach as compared to training from scratch. Among the popular variants of convolutional neural networks (CNN), the study has used ResNet50, owing to its proven classification potential as reported in the literature [1, 2]. The classification result for two-class classification is 98.6% and for seven-class classification the classification accuracy is 88.5% with the most abnormal class (Squamous Cell Carcinoma (SCC)) showing a good classification rate of 91.8% and precision value of 0.907. The classification accuracy for slide-level classification is 78.09%. The results show the tremendous potential of the Deep neural Networks trained through transfer learning for the classification of uterine cervix cancer cases both at the cell level (2 classes, 7 classes) and slide level.