Genetic Algorithm-Based Weighted Ensemble of Deep CNN Models for Dental Disease Detection
Abstract
Accurate categorization of dental imagery is crucial to facilitate early identification and efficient treatment strategizing in dental healthcare. This investigation aspires to establish a multiclass dental pathology classification framework employing a Genetic Algorithm (GA)-based weighted ensemble of deep Convolutional Neural Network (CNN) architectures. The resources used in this scholarly inquiry comprise a publicly accessible dental image repository encompassing four categories: caries, calculus, discoloration, and hypodontia, sourced from diverse clinical and web-based sources. The proposed methodology implements transfer learning by fine-tuning three pre-trained CNN frameworks, specifically ResNet50V2, EfficientNetB4, and EfficientNetB5. Each foundational model is trained autonomously, and their predictive probabilities are amalgamated via a weighted ensemble approach. A Genetic Algorithm is utilized to autonomously ascertain the optimal ensemble weights by maximizing the macro F1-score, thereby ensuring equitable performance across all pathological categories. Experiments were executed utilizing varying training epochs to assess model convergence and generalizability. The experimental findings reveal that the GA-based ensemble attained optimal efficacy at a configuration with a maximum of 300 training epochs, with an accuracy of 98.71%, a precision of 98.62%, a recall of 97.66%, and an F1-score of 98.12. A minor reduction in performance at elevated epochs indicates the emergence of overfitting. Compared with individual CNN models and non-optimized ensemble methods, the proposed framework exhibits enhanced robustness and more consistent class-wise classification. These results underscore the efficacy of metaheuristic optimization in advancing ensemble deep learning for dental disease classification. Prospective research should concentrate on validation utilizing larger multicenter clinical datasets, exploration of alternative optimization algorithms, and real-time clinical implementation.