Artificial Intelligence Techniques for Automatic Cervical Cancer Detection and Segmentation Using Sparsity-Aware Orthogonal Initialization in Deep Neural Network Classifiers: Trends and Challenges

Abstract

Cervical cancer remains a significant global health concern, particularly in low-resource regions where access to timely screening and expert diagnosis is limited. The integration of artificial intelligence (AI), especially deep learning techniques, has greatly enhanced the automation of cervical cancer detection and segmentation from medical images such as Pap smear slides, colposcopy images, and histopathological data. Deep learning architectures including convolutional neural networks (CNNs), U-Net variants, and hybrid CNN-transformer models have demonstrated high accuracy in classification and segmentation tasks by effectively capturing complex spatial and hierarchical features. However, training these deep models presents challenges such as vanishing gradients, overfitting, and slow convergence. Sparsity-aware orthogonal initialization has emerged as an effective optimization approach, improving training stability, preserving signal propagation, and reducing redundancy in network parameters. This review highlights recent advancements in segmentation techniques, hybrid architectures, and optimization strategies, along with the use of attention mechanisms, transfer learning, and data augmentation to enhance performance. Despite progress, challenges such as limited annotated data, class imbalance, interpretability, and computational demands persist, indicating the need for more efficient and explainable AI-based diagnostic systems.