Optimised deep learning for oral cancer classification

Abstract

Oral cancer detection is essential, especially in areas with high occurrence rates, is essential for better early diagnosis and individualised treatment plans. SV-OnionNet is a deep learning framework presented in this article that aims to improve the classification accuracy of oral cancer diagnosis. While maintaining important structural details, the method lowers noise in medical pictures by integrating an adaptive Non-Linear Means (NLM) filter. Spatial features are improved by the Label-Guided Attention (LGA) module, which guarantees constant labelling and improves feature extraction. By enabling accurate pixel-level segmentation of lesions, Seg-UNet provides increased classification reliability. The Support Vector Machines (SVM) deep learning classification model used in the SV-OnionNet architecture preserves spatial relationships for improved feature learning, replacing traditional fully linked layers (LKN). The Competitive Search Optimization (CSO) algorithm fine-tunes model parameters, therefore optimising feature selection and classification. The evaluation on the Mouth and Oral Diseases dataset demonstrated exceptional accuracy, precision, recall, and specificity, with the proposed classification achieving a 99.94% accuracy. These findings emphasise the effectiveness of SV-OnionNet in improving the diagnostic accuracy and reliability. The study highlights the potential of integrating deep learning techniques with optimisation strategies to advance oral cancer detection. Future research will focus on expanding datasets and exploring additional optimisation methods to further improve the classification performance.