Deep Learning Approach to Cervical Cancer Classification

Background: Cervical cancer is caused in the vast majority of cases by the human papillomavirus through sexual contact and requires a specific analysis with specialised equipment and health professionals to be detected. However, its incidence is up to ten times higher in areas without such resources. To overcome these shortcomings and perform mass screening, classifiers based on convolutional neural networks are being developed and, to speed up sample collection, liquid cytology samples have been used in recent years. Even so, these systems are not infallible and require verification by a healthcare professional to make a final diagnosis, which is why explainable deep learning techniques are taking advantage in this area.Objective: In this work, we perform an exhaustive analysis of the classifier systems developed in recent years and develop our own classifier through the evaluation of multiple combinations and the detailed adjustment of the system's hyperparameters.Methods: The classifiers developed (for two and four classes following the Bethesda System nomenclature) are evaluated in depth and compared with the works published in the last five years in terms of both efficiency and computational cost. Moreover, a complementary report generation system is presented which, in addition to the classifier results, provides the healthcare professional a detailed report containing the classification confidence of each of the categories and the visual annotation on the original image of the area (or areas) where the classifier has been focused to obtain its results.Results: The results determine an accuracy close to 98\% for the four-class classifier (with no false positives for the negative class) and 100\% for the two-class classifier. The final report include a detailed confidence list and a heat map containing the regions of interest used by the system to classify the given image.Conclusions: Compared to the previous works, this classifier obtains better accuracy results with lower computational cost. Moreover, the use of Explainable Artificial Intelligence techniques is an useful tool for the healthcare professional, as it determines the confidence in the results and the areas of the image used for classification, which can be used for an incremental learning process (allowing the professional to point out the areas he/she thinks are most important and cross-checking them against those detected by the system).