# A novel Pulp Caries GAN multi loss GAN with new pulp inspired metaheuristics for pediatric dental caries detection and segmentation

## TL;DR

This paper introduces a new GAN called Pulp-Caries-GAN that generates realistic dental images to improve early detection of caries in children.

## Contribution

A novel GAN with a pulp-inspired metaheuristic and multi-loss architecture for pediatric dental caries detection and segmentation.

## Key findings

- Pulp-Caries-GAN achieved high-quality synthetic image generation with FID of 154.87 and Inception Score of 80.12.
- Synthetic images improved segmentation performance, with a Hierarchical Dense U-Net achieving 95.12% Dice coefficient.
- 87% of synthetic images were clinically indistinguishable from real radiographs according to pediatric dentists.

## Abstract

Early detection of dental caries in pediatric populations remains challenging due to limited annotated datasets and the subtle manifestation of incipient lesions. This study introduces Pulp-Caries-GAN, a novel generative adversarial network incorporating a biomimetic optimization strategy for high-fidelity synthetic dental image generation. The framework integrates a multi-loss architecture combining adversarial, pixel-wise, perceptual, and structural similarity losses with a pulp-inspired metaheuristic function that models neurophysiological dynamics of dental pulp tissue to preserve anatomical coherence. The optimization strategy employs spatially-adaptive regularization through an anatomical masking mechanism that enforces tissue-specific constraints based on diagnostic importance. Experimental validation was conducted on a pediatric dental panoramic dataset comprising 193 annotated images from 106 patients aged 2–13 years. The results demonstrate superior image synthesis quality compared to conventional GAN architectures, achieving a Fréchet Inception Distance of 154.87, Inception Score of 80.12, and Peak Signal-to-Noise Ratio of 80.04. Integration of synthetic images generated by Pulp-Caries-GAN significantly enhanced segmentation performance across multiple U-Net variants. The Hierarchical Dense U-Net achieved optimal results with a Dice coefficient of 95.12%, accuracy of 95.65%, precision of 95.32%, and recall of 93.7%. Ablation studies confirmed the critical role of the pulp-inspired loss component and anatomical masking in maintaining structural integrity while reducing artifacts in synthetic images. Clinical validation by five board-certified pediatric dentists revealed that 87% of synthetic images were clinically indistinguishable from real radiographs, with 94% of synthetic lesions exhibiting anatomically correct progression patterns. These findings demonstrate the efficacy of biomimetic optimization approaches in medical image synthesis and establish a robust framework for automated pediatric dental caries detection with potential for clinical translation.

The online version contains supplementary material available at 10.1038/s41598-025-28459-8.