Xception U-Net with augmentation and preprocessing achieves Dice 0.95 for dental caries segmentation

An Xception-based U-Net with data augmentation and preprocessing achieved a Dice coefficient of 0.9517 for dental caries segmentation on panoramic radiographs, outperforming standard U-Net (Dice 0.7380) and models without augmentation, which failed dramatically (Dice as low as 0…

• Comparative study on 500 manually annotated panoramic dental radiographs, using five-fold cross-validation with four U-Net backbones (standard, VGG16, ResNet50, Xception).
• Without augmentation and preprocessing, model performance collapsed: ResNet50 achieved Dice 0.0068, demonstrating the critical role of data preparation.
• Limitation: Single-institution dataset, internal cross-validation only, no external or prospective clinical validation.

RadPigeon summaries are original and for information only. They are not clinical advice.