Impact of image processing techniques on deep learning-based classification accuracy of cervical vertebral maturation.

objectivesThis research aimed to investigate how different image processing techniques influence the performance of a deep learning model in classifying cervical vertebral maturation (CVM) stages.

methodsA dataset of 799 cephalometric radiographs, originally obtained for orthodontic diagnosis, was analyzed. For each image, a rectangular section including the second to fourth cervical vertebrae (C2-C4) was extracted and used as the standard image. A version with low signal intensity was created and termed the "low-density image." Additionally, images were labeled by filling the internal areas of C2, C3, and C4 with yellow, generating three labeled variants: standard labeling, low-density labeling, and dual-color labeling. All image types were classified into six CVM stages (CS1-CS6) using a convolutional neural network based on the AlexNet architecture, implemented via Neural Network Console. From the dataset, 641 images were allocated for learning and 158 for testing. Classification accuracy was evaluated across all image types.

resultsThe average classification accuracies for the standard and low-density images were 46.8% and 48.7%, respectively. In contrast, labeled images performed significantly better, with accuracies of 82.3% (standard labeling), 81.0% (low-density labeling), and 88.0% (dual-color labeling).

conclusionsAmong the image variations, the dual-color labeling dataset achieved the highest classification performance. These results suggest that enhanced labeling techniques can significantly improve the reliability of deep learning models for automated assessment of skeletal maturity.