Oral cancer detection using feature-level fusion and novel self-attention mechanisms.

The rising prevalence of oral and dental conditions, including issues like gum disease and oral cancer, presents a pressing global health challenge. The promptly identification of individuals at risk of developing these potentially life-threatening oral ailments, especially those linked to lip and oral cavity cancer. Early detection is crucial for effective intervention and containment of these diseases. Several studies in medical research have explored the use of convolutional neural network (CNN) models with pre-trained weights to identify oral and dental diseases. Previous research has primarily focused on amalgamating predictions from various models to boost accuracy, the integration of diverse model predictions presents challenges such as potential inconsistencies, conflicting outcomes, and heightened model complexity. However, utilizing a variety of model features can not only boost the robustness and accuracy of the detection process but also streamline the overall complexity of the model ensemble. The challenge lies in finding more effective methods to combine and leverage features extracted from deep learning models for superior predictive accuracy. To tackle this challenge, our research introduces an innovative approach that incorporates feature-level fusion techniques for oral cancer detection. We have developed a novel self-attention block as a foundational element of our approach. Additionally, we utilize transfer learning (TL) models for feature fusion, specifically EfficientNetB0 and EfficientNetB1, renowned for their effectiveness in image classification tasks. During the training process, we incorporate a self-attention block, a crucial innovation that facilitates the fusion of features extracted from these EfficientNet models. Our proposed methodology revolves around the concept of feature-level fusion. We employ a novel self-attention block and leverage transfer learning models to extract and combine relevant features. This fusion strategy enables us to collective strengths of these models, resulting in more accurate predictions than what can be achieved using each model in isolation. To evaluate the effectiveness of our approach, we conducted rigorous assessments using the well-established MOD dataset, which is a publicly available benchmark in the field of oral cancer detection. Our model achieved an impressive accuracy rate of 98.83% in these evaluations. [ABSTRACT FROM AUTHOR]