DentoSeg-SSL | M N Aditya

Dental radiograph segmentation is a prerequisite for automated detection of cavities, fractures, and periodontal disease, yet large labelled datasets are expensive to obtain. DentoSeg-SSL addresses this by combining self-supervised contrastive pre-training with a U-Net segmentation head, enabling the model to extract robust features from unlabelled panoramic radiographs before fine-tuning on a small annotated set.

Grad-CAM heat-maps are generated post-hoc to highlight the regions driving each segmentation decision, bridging the gap between clinical trust and AI automation.

Key contributions

Contrastive pre-training on unlabelled panoramic X-rays reduces annotation dependency.
Attention-based U-Net decoder achieves accurate boundary delineation of teeth and surrounding structures.
Grad-CAM integration provides per-prediction visual explanations accepted by clinical reviewers.

Tech stack: Python · PyTorch · U-Net · Grad-CAM · JSCP 2025 (Sreeram et al., 2025)

References

2025

J
Explainable AI for Panoramic Dental Radiographs using Contrastive Learning and UNet Based Segmentation

Aadyoth Sreeram, R. Balamurugan, and M. N. Aditya

Journal of Soft Computing Paradigm, 2025

Abs Bib HTML

Self-supervised segmentation of panoramic dental radiographs often lacks interpretability, leading clinicians to be hesitant in relying on model predictions. We propose a hybrid strategy that integrates attention-based U-Net segmentation with contrastive self-supervised pre-training, enabling robust feature extraction from unlabelled radiographs before supervised fine-tuning. Grad-CAM heat-maps are generated post-hoc to highlight regions driving each segmentation decision, bridging the gap between clinical trust and AI automation. The approach detects anomalies including cavities, fractures, and periodontal disease with high accuracy while providing visual explanations that improve diagnostic confidence.
@article{aditya2025dental, title = {Explainable AI for Panoramic Dental Radiographs using Contrastive Learning and UNet Based Segmentation}, author = {Sreeram, Aadyoth and Balamurugan, R. and Aditya, M. N.}, journal = {Journal of Soft Computing Paradigm}, volume = {7}, number = {2}, pages = {114--123}, year = {2025}, }