When CNNs Outperform Transformers and Mambas: Revisiting Deep Architectures for Dental Caries Segmentation

TL;DR

This study tackles automated segmentation of dental caries in panoramic radiographs by benchmarking 12 architectures across CNNs, vision transformers, and Mamba models on the DC1000 dataset. Using a unified training pipeline and consistent metrics, it finds that CNN-based DoubleU-Net achieves the best performance (mIoU ≈ 0.598, mDSC ≈ 0.735) and that transformers and Mamba models, despite strong discrimination (AUC > 0.87), struggle with precise boundary delineation in data-limited settings. The results emphasize the importance of spatial inductive priors and architecture-task alignment over model complexity in medical image segmentation. The work provides a practical reference for clinical deployment and highlights the need for larger, multi-institutional datasets to generalize beyond DC1000 and extend to multi-class dental pathologies.

Abstract

Accurate identification and segmentation of dental caries in panoramic radiographs are critical for early diagnosis and effective treatment planning. Automated segmentation remains challenging due to low lesion contrast, morphological variability, and limited annotated data. In this study, we present the first comprehensive benchmarking of convolutional neural networks, vision transformers and state-space mamba architectures for automated dental caries segmentation on panoramic radiographs through a DC1000 dataset. Twelve state-of-the-art architectures, including VMUnet, MambaUNet, VMUNetv2, RMAMamba-S, TransNetR, PVTFormer, DoubleU-Net, and ResUNet++, were trained under identical configurations. Results reveal that, contrary to the growing trend toward complex attention based architectures, the CNN-based DoubleU-Net achieved the highest dice coefficient of 0.7345, mIoU of 0.5978, and precision of 0.8145, outperforming all transformer and Mamba variants. In the study, the top 3 results across all performance metrics were achieved by CNN-based architectures. Here, Mamba and transformer-based methods, despite their theoretical advantage in global context modeling, underperformed due to limited data and weaker spatial priors. These findings underscore the importance of architecture-task alignment in domain-specific medical image segmentation more than model complexity. Our code is available at: https://github.com/JunZengz/dental-caries-segmentation.

Figures (4)

• Figure 1: Variation in dental panoramic X-ray structures across the DC1000 dataset, showing patients with full dentition, missing teeth, and braces.
• Figure 2: Illustration of dental caries annotation in the DC1000 dataset. The figure shows the original panoramic X-ray image, the expert-annotated binary mask highlighting carious regions, and the overlay visualization combining both.
• Figure 3: Qualitative comparison of segmentation outputs from four representative architectures on the DC1000 test set. The first column shows the ground truth masks, followed by predictions from DoubleU-Net, U-Net, PVTFormer, and RMAMamba-S. DoubleU-Net consistently yields the most accurate lesion outlines, preserving fine radiolucent boundaries and diffuse enamel–dentin transitions that other models often miss.
• Figure 4: ROC curves comparing pixel-level discriminative performance of several CNN, Transformer, and Mamba architectures.