Deep Learning-Based Bone Age Assessment from Hand X-Rays: An Evaluation and Analysis

In the realm of deep learning-based bone age assessment from hand X-ray scans, this study thoroughly evaluates several convolutional neural network (CNN) architectures: VGG19, ResNet50, AlexNet, SENet, and PyramidalNet. Utilizing a substantial dataset generously provided by the Radiological Society of North America (RSNA), with 12,611 images for training and 200 for testing, this research explores the intricacies of medical image analysis. Among the models, PyramidalNet stands out with the highest R2 score of 0.74, demonstrating exceptional performance in explaining variance and a low Mean Absolute Error (MAE) of 2.943, indicating precise bone age predictions. AlexNet also excels, boasting a high R2 score of 0.66 and a low MAE of 2.76. While VGG19 achieved a commendable R2 score of 0.71, it had a higher MAE of 17.46. SENet, with a similar R2 score of 0.64, exhibited a slightly elevated MAE of 4.113. In contrast, ResNet50 fell behind with least R2 score of 0.35 and highest MAE of 25.73. This comprehensive analysis underscores the importance of selecting the most suitable deep learning architecture for precise bone age assessment, positioning PyramidalNet and AlexNet as robust choices for accurate predictions, and advancing the field of medical image analysis.