Cue-Aided Multi-Class Segmentation (CAMS): A Novel Approach for Enhanced Modelling in Dental Panoramic Imaging

Abstract

Effective segmentation in medical imaging is essential for accurate diagnosis, particularly in complex multi-class scenarios like dental panoramic segmentation. Traditional methods often struggle with class imbalance and overlapping regions. This paper introduces Cue-Aided Multi-Class Segmentation (CAMS), a novel method that enhances class discrimination by using an additional cue channel. Unlike conventional approaches that train separate models for each class or use a single model for all classes, CAMS processes each sample multiple times with a unique mask for one class, encoded through fixed image regions corresponding to dental quadrants. We utilize the UNet architecture for its simplicity and proven effectiveness in medical segmentation. CAMS was validated using the DENTEX dental panoramic dataset and compared against two traditional methods: a Specialized model for each class and an All-In-One model. CAMS achieved an Intersection over Union (IoU) of 0.832 and a Dice coefficient of 0.904, outperforming the All-In-One model, which had a mean IoU of 0.684 and a Dice coefficient of 0.737. The Specialized model achieved a similar IoU of 0.835 and a Dice coefficient of 0.907. Preliminary results suggest that CAMS offers comparable accuracy to existing methods but with significant computational efficiency, making it a promising approach for multi-class medical imaging tasks.