Advancements in fusion-based deep representation learning for enhanced cervical precancerous lesion classification...

One such prevalent kind of cancer among women is cervical cancer (CC). Fatality rates and incidence are progressively increasing, mainly in developing countries, due to a lack of experienced specialists, inadequate public awareness, and limited screening facilities. Nevertheless, CC cells exhibit composite textural features, and smaller changes among dissimilar cell subcategories result in greater challenges for the higher-accuracy screening of CC. This systematic analysis aims to assess the predictive value of artificial intelligence (AI) technologies for diagnosing, screening, and predicting CC and precancerous lesions. Deep learning (DL) and AI generally have a positive impact on computer-aided clinical diagnosis, particularly with the increasing accessibility of larger amounts of medical data that can aid AI methods in achieving high performance on various medical tasks. In this paper, a Fusion of Advanced Feature Reduction and Deep Representation Learning Approaches for Cervical Precancerous Lesion Classification (FAFRDRL-CPLC) technique using biomedical image analysis is proposed. The primary purpose of the FAFRDRL-CPLC technique is to serve as a valuable tool for assisting clinicians in the initial study and treatment planning of cervical precancerous lesions. Initially, the FAFRDRL-CPLC approach applies an anisotropic diffusion filtering (ADF) method for pre-processing to reduce noise while preserving crucial edges and lesion details. Furthermore, the fusion of advanced feature reduction models, such as the maximally scalable vision transformer (MaxViT-v2), the simple framework for contrastive learning of visual representations (SimCLR), and the Twins-spatially separable vision transformer (Twins-SVT) models, is employed to capture diverse and complementary representations from the pre-processed images. Finally, the stacked auto-encoder (SAE) classifier is utilized for the precancerous lesion detection process. The FAFRDRL-CPLC method is examined through experimentation using the Malhari dataset. The comparison study of the FAFRDRL-CPLC method demonstrated a superior accuracy value of 98.62% over existing approaches.