Design of a novel panoptic segmentation using multi-scale pooling model for tooth segmentation

Utilizing individual annotation of panoramic radiographs, a comprehensive deep learning multi-scale spatial pooling (ms-SP)-based panoptic segmentation technique is tested for its effectiveness in segmenting teeth autonomously. On a panoramic radiograph, each tooth was meticulously tagged by an oral radiologist to accurately depict its real structure. From the initial data points, we used the augmentation strategy to create training samples to reduce over-fitting. With the proposed multi-scale spatial pooling (ms-SP), a completely deep learning approach was used to locate and identify the dental traits. Performance was evaluated using the F1 score, and visual analysis. The suggested method resulted in a mean IoU of 87% and an F1 score of 98.9%, accuracy of 98.5%, recall of 93%, precision of 94.5%, dice score of 94.5% and PFOM is 80.5%. The segmentation technique was evaluated visually, and the results were very similar to the actual data. The technique produced effective results for automating the segmentation of teeth on panoramic dental photos. The suggested technique may be advantageous for the first stages of forensic identification and diagnostic automation, which both involve similar segmentation tasks.