SFI-ensemble: Sugeno fuzzy integral-based ensemble of CNN models with meta-heuristic fuzzy measures for mouth and oral disease detection

The rising prevalence of mouth and oral diseases (MOD), including gum disease and oral cancer, presents a major global health challenge, where timely detection is crucial for effective intervention. Recognizing the limitations of a single learning model in capturing intricate information for precise disease prediction from complex data, we introduce a robust deep learning ensemble framework named Sugeno Fuzzy Integral (SFI)-Ensemble in this paper. Our methodology involves a meticulous preprocessing of the dataset utilizing fuzzy contrast enhancement (FCE) to enhance data quality and contrast. Additionally, we propose a reconstruction approach that employs transfer learning (TL) and fine-tuning on four Convolutional Neural Network (CNN) models—DenseNet121, MobileNetV1, DenseNet169, and Resnet101V2—to optimize their architectures specifically for MOD classification. The focal point of this contribution lies in the introduction of a groundbreaking ensemble method. This ensemble method dynamically combines decision scores from the CNN models using the SFI-based technique, offering a resilient and adaptive approach that factors in the confidence of base learners’ predictions through fuzzy integrals. To overcome the prevalent challenge of experimentally defining fuzzy measures in ensemble methods based on fuzzy integrals, we surpass conventional manual tuning. Our approach involves the utilization of seven distinct meta-heuristic optimization algorithms for the optimal determination of fuzzy measures. This not only ensures stability but also highlights the effectiveness of the proposed SFI-Ensemble. A comprehensive assessment is carried out on publicly accessible datasets to detect MOD, complemented by Grad-CAM interpretability and meticulous statistical analyses. Additionally, we benchmark the results against baseline models and state-of-the-art methods, with our proposed framework consistently surpassing them, attaining an impressive accuracy of 99.70%. This underscores the superior performance and robustness of our proposed methodology in contrast to traditional ensemble methods. Our approach, integrating dataset preprocessing, model reconstruction, and ensemble innovation, provides doctors with an effective tool for accurate MOD diagnosis, enhancing adaptability and performance through fuzzy integral-based fusion.