A Multimodal Approach for Pap-Smear Image-Based Classification of Cervical Cancer

Among the most prevalent malignancies in women cervical cancer is one of them. It is challenging to detect cervical cancer early. Pap-smear images are manually analyzed by skilled pathologists to detect cervical cancer. Although it is a difficult and time-consuming process, the use of pap smear images for cervical cancer treatment depends on an accurate and fast diagnosis. Furthermore, although it is clinically significant, classifying cervical cells into discrete groups is challenging in the cervical cancer diagnosis. Deep learning-based algorithms are used to construct computer-aided diagnostic (CAD) systems, which offer a viable solution. For automatic distinguishing of cervical cancer cells from pap-smear imagery, this study proposes a multimodal approach integrating the hand-crafted cellular feature information, a convolutional neural network ResNet-18 and graph convolutional network (GCN). Visual characteristics are retrieved via ResNet-18, and GCN captures the spatial relations of superpixel-based graph representations. By combining these modalities and using a cross-modal attention mechanism, the model can ascertain the relative significance of each for accurate classification. The proposed method achieves 100% recall, 95.21% accuracy and AUROC of 0.9969, surpassing current approaches with the benchmark dataset, SIPaKMeD with 4049 samples and its corresponding hand-crafted features.