This paper presents an extended comparative evaluation of Mask R-CNN and YOLOv8 for automatic tooth counting and FDI numbering in panoramic radiographs. The study uses a private clinical dataset with 410 anonymized images and 53 dental classes, covering permanent teeth, deciduous teeth, and a supranumerary class. The manuscript preserves the original inter-architecture comparison between a two-stage Mask R-CNN reference and one-stage YOLO-based segmentation, while extending the reproducible experimental evidence on the YOLOv8 branch through model-scale comparison, expanded augmentation, calibrated post-processing, class-group analysis, five-fold robustness assessment, and a countoriented audit. On the fixed validation split (328 training images, 82 validation images), YOLOv8s-seg achieved the best mAP50 for mask segmentation (0.9003), while YOLOv8m-seg achieved the best mAP50–95 (0.6673). A five-fold validation with the medium protocol produced mean scores of 0.8770 ± 0.0132 for mAP50 and 0.6374 ± 0.0165 for mAP50–95. The count-oriented audit of the post-processed medium configuration yielded a mean absolute counting error of 2.49 teeth per image, exact count agreement in 23.2% of validation cases, and dentition-type agreement in 81 of 82 images. The results preserve the original comparative message while strengthening the evidence that the small and medium YOLOv8 configurations provide the strongest practical trade-off, whereas mixed dentition, rare classes, and exact tooth counts remain the main open difficulties.