Abstract
The prediction of Cervical Cancer (CC) remains a tough task due to diverse clinical variations and unbalanced data distribution, while good-quality data remains limited. Early CC signs tend to lack distinct characteristics, which makes their precise identification more challenging for medical professionals. Different methods of diagnosis integrated with the unique attributes of different datasets make it difficult to apply a uniform model to generalized patient populations. To address these issues, this paper introduces the Deformable Kernel Darknet-53 with Depth-Wise Separable Convolutional Neural Network (DK-D53-DWSCNNet). This framework initiates with image enhancement and quality adaptation for diverse heterogeneous morphological structures using Deformable Kernel Networks for Joint Image Filtering (DKNet-JIF). An ensemble Darknet-53 Convolutional Neural Network with a Contextual Attention Network (D53-CNN-CAN) is used to perform segmentation, which improves the representation of variable lesion patterns. A hybrid GAT-DWSCNNet, which merges geometric algebra transformers (GAT) with depth-wise separable CNNs, is used for feature extraction and classification, allowing the system to capture both edge and contextual features while reducing redundancy. The Hyperbolic Sine Optimizer (HSO) optimizes model training by providing the best trade-off between convergence speed and accuracy. With an accuracy of 99.9% and sensitivity of 99.8%, the experiments on the Herlev and SEER datasets depict remarkable generalization. These findings demonstrate the potential of DK-D53-DWSCNNet in improving early and robust CC prediction across different datasets, supporting the model to be embedded in clinical diagnostic procedures.