SwinTransConv and Tab-FCNN: a novel SwinTransConv neural network features and tab-fully connected neural network in pap smear images for cervical cancer classification

The primary objective of this paper is to delineate a Deep Learning (DL) methodology for cervical cancer from Pap smear imagery. In the interest of augmenting the quality and equilibrium of the dataset, the initial phase involved executing ROI detection on the Pap smear images. ROI detection is executed using YOLO V4 model in the input Pap smear images for detecting superficial, intermediate and parabasal layers. Then, the segmentation phase is performed to delinate cytoplasm and the nucleus, employing the YOLOv11 model. Subsequently, the feature extraction is executed by the proposed SwinTransConv, which integrates the Swin Transformer with a Convolutional Neural Network (CNN) to yield a robust and hierarchical representation of salient cellular features. The derived features act as input for the classification phase, for which a Tabular-Fully Convolutional Neural Network (Tab-FCNN) model is proposed by combining Tabular Network (TabNet) and Fully Convolutional Neural Networks (FCNN). TabNet identifies significant features from the input dataset utilizing attention-based mechanisms tailored for tabular data, whereas FCNN enhances the final decision-making process by assimilating complex feature interactions. Experimental findings state that the proposed approach reached an accuracy of 97.9%, a sensitivity of 95.4% and a specificity of 99.4%.