An accurate pediatric bone age prediction model using deep learning and contrast conversion

<strong>Purpose:</strong> The purpose of this study was to develop an accurate pediatric bone age prediction model using deep learning models and contrast conversion techniques, with the goal of improving growth assessment and clinical decision-making in clinical practice. <strong>Materials and Methods:</strong> The study utilized various deep learning models and contrast conversion techniques for bone age prediction. Training data comprised pediatric left-hand X-ray images with bone age and gender information. Models like CNN, ResNet 50, VGG 19, Inception V3, and Xception were trained and evaluated using MAE. Contrast conversion techniques, including FCE, CLAHE, and HE, were applied to test data, and image quality was assessed using PSNR, MSE, SNR, COV and CNR metrics. The bone age prediction results using the test data were evaluated based on MAE and RMSE, and a t-test was performed. <strong>Results:</strong> The Xception showed the best performance with an MAE of 41.12, outperforming CNN, ResNet50, VGG 19, and Inception V3. HE exhibited superior image quality with higher SNR and COV values compared to other methods. Additionally, HE demonstrated the highest contrast among the techniques assessed, with a CNR value of 1.29. Enhancements in bone age prediction resulted in a decline in MAE between 0.24 and 2.11, along with a decrease in RMSE ranging from 0.02 to 0.21. <strong>Conclusion:</strong> This study shows that without preprocessing the data before model learning, there is not a significant difference in bone age prediction performance between contrast-converted images and original images.