Cervical Cancer Screening using Explainability

Cervical cancer is a significant global health issue, and traditional screening methods like Pap smears are labor intensive and may miss cases. Cervical cancer, which is ranked fourth among cancers affecting women, is highly treatable when detected early through the pap smear test. Deep Learning (DL) models, particularly Convolutional Neural Networks (CNNs), analyze pap smear images, yet their ``Black-Box" nature raises transparency concerns in medical diagnostics. Automation is needed, but it faces challenges in interpretability and data availability.  This study introduces a solution named EnsembleCAM to enhance interpretability by unifying visual explanations through the combination of diverse Class Activation Maps (CAMs). We use Explainable Artificial Intelligence (XAI) techniques like GradCAM, GradCAM++, and LRP to improve the transparency and interpretability of a cervical cell classification model, making it a novel contribution to enhancing the trustworthiness of automated cervical cancer detection.

Using the Herlev Dataset, we employ data pre-processing, data augmentation techniques, develop an XceptionNet based binary classification model with an accuracy of 89\% and apply GradCAM, GradCAM++, Score-CAM, Eigen-CAM and LayerCAM on this classifier.  Through VGG16, an accuracy of 91.94% was achieved in classifying cervical cancer cells. The qualitative analysis of XAI methods confirmed that the model relied on nucleus and cytoplasm features, key indicators of malignancy. With the least mean image entropy of 2.4849 and steep prediction confidence drop with perturbations, quantitatively proved Layer-wise Relevance Propagation (LRP) to be the most effective XAI technique for cervical cell classification.

Then, the novel EnsembleCAM is constructed taking the median of activation maps from the five individual CAM methods. The analysis of activation maps of each CAM method and EnsembleCAM confirmed that in activation maps of EnsembleCAM, higher activation values were more concentrated around the nucleus which is the most important region in indicating cervical malignancy. The evaluation using pixel flipping performance metric also proved that the EnsembleCAM effectively recognises regions vital to the model’s decision-making through its steepest drop in the mean prediction score when the pixels in the region contributing most to the model’s decision were flipped.