Diagnosis of Cervical Cancer Using a Deep Learning Explainable Fusion Model

Cervical cancer continues to be a significant global health issue, ranking as the fourth most prevalent cancer affecting women. Enhancing population screening programs by refining the examination of cervical samples conducted by skilled pathologists offers a compelling alternative for early detection of this disease. Deep Learning facilitates the development of automatic classification models to aid experts in this task. However, it is increasingly important to bring explainability to the model in order to understand how the network learns to identify pathology. In this paper, the explainability created by a heatmap, using the Gradient-weighted Class Activation Mapping (Grad-CAM) technique, is merged with the original image by studying the different intensities for the overlap by means of a hybrid architecture composed by a Convolutional Neural Network and explainability techniques. Through this blending, a new image of the cell is created for training where the heatmap provides the original image of the cell with information about the location of the region of interest. Finally, it is observed that a 10% intensity provided by the heatmap is the most efficient value for this fusion, reaching accuracy values of 94% in a model that indicates whether or not a revision by the pathologist is necessary.