Feature-Level Fusion-Based CNN-Transformer-KAN Hybrid Approach for Dental Implant Diameter and Length Detection

Accurate and automated determination of implant diameter and length is a significant but challenging image classification problem for dental planning and clinical decision support systems. In this study, a hybrid model was developed for implant diameter and length determination. First, different Convolutional Neural Network (CNN)-based backbone models were fine-tuned to identify the most successful models in terms of performance. Then, unlike classical ensemble approaches, the models yielding the most successful results were combined at the feature level, not the decision level, to create a dual-backbone hybrid structure. To model canal dependencies, the Efficient Channel Attention (ECA) mechanism was integrated into the proposed hybrid model, and then a Transformer encoder-based attention structure was added to the proposed model for learning global contextual relationships. In addition, a Kolmogorov-Arnold Network (KAN)-based classifier was preferred instead of traditional linear structures in the classification layer to ensure more effective learning of nonlinear decision boundaries. A comprehensive ablation study was performed to analyze the contribution of the model components. The proposed model was compared with different pre-trained models. In conclusion, the proposed model was observed to be more successful than other models with a test accuracy of 88.46%. Based on this accuracy value, we believe the proposed model can contribute to the development of automated analysis and decision support systems in the field of dental imaging.