An Intelligent Multimodal Deep Learning with Incremental and Bio-Inspired Optimization for Early Oral Cancer Detection

Abstract

Oral cancer poses a major public health challenge around the globe, and when diagnosed late mortality rates are high. Early detection remains critical in improving patient outcomes and rates of survival. As far as could be judged, this is the first of its kind of study to provide a complete diagnostic multimodal deep learning intelligent system that incorporates CNNs and RNNs with state-of-art transfer learning methodology for early oral cancer detection. To allow continuous model adaptation over a wide spectrum of clinical scenarios, our proposed framework integrates incremental learning methods with bio-inspired optimization techniques based on Deep Q-Networks (DQN). Our model is reinforced with state-of-the-art datasets such as ORCHID, MODID along with the newly crowdsourced multimodal collections which deal with such a major shortcoming identified in public oral cancer data. Our performance evaluation highlights the robustness of model generalisation over extant methods, with particularly good results observed through incorporating attention mechanisms and transformer-based architectures. The proposed system demonstrates robust performance in diverse patient populations while preserving clinical interpretability and real-time processing speed