QEEF: A Quantitative Explainability Evaluation Framework for CNN and Vision Transformer-Based Segmentation Models in Dental Images

Deep learning methods have exemplified performance in various dental image segmentation tasks. However, their interpretability inadequacy remains a hindrance to clinical adoption. This study introduces a quantitative explainability-driven evaluation framework to systematically assess and compare convolutional and transformer-based segmentation models in dental imaging. Specifically, we evaluate a convolutional Watershed Encoder–Decoder Neural Network (WEDN) and the hybrid Vision Transformer U-Net (ViTUNet) model using both qualitative and quantitative explainability metrics. The qualitative analysis employs Grad-CAM, occlusion sensitivity, and attribution-based visualization techniques to assess model decision patterns. Quantitatively, we define and apply explainability metrics, including Dice coefficient, boundary Dice, sparsity, and faithfulness metric scores, to independently evaluate segmentation performance and explanation quality. The experimental results show that ViTUNet achieves superior spatial localization, with higher Dice (0.765) and boundary Dice (0.421), compared to the WEDN model (0.664 and 0.221), indicating improved lesion delineation in dental images. Moreover, the ViTUNet model exhibits higher sparsity scores in attribution maps, indicating more concentrated and structured interpretability regions, while the WEDN model demonstrates higher faithfulness scores under perturbation-based evaluations, thus reflecting stronger dependence on localized feature representations. These findings highlight a key distinction between convolutional and transformer-based architectures with predictive behavior and explanation characteristics. Notably, the results further indicate that gradient-based explanation methods are more stable for transformer-based models, while perturbation-based methods better capture the decision logic of convolutional networks. In general, this study establishes an integrated framework for collectively evaluating segmentation performance and explainability, allowing more transparent and clinically reliable deployment of deep learning models in dental imaging.