Detecção de Maturidade Óssea utilizando algoritmo YOLO e Redes Neurais Convolucionais

Abstract

Bone age assessment is an essential diagnostic tool in pediatric and endocrinological practice, used to monitor skeletal maturation and assist in the diagnosis of growth disorders. Traditional assessment methods, such as the Greulich and Pyle (GP) atlas or the Tanner-Whitehouse (TW) scoring system, rely on the visual analysis of hand and wrist radiographs. However, these manual approaches are known to be time-consuming, subjective, and prone to considerable inter- and intra-observer variability, which may compromise the consistency of clinical follow-up. Machine learning emerges as a solution to reduce these limitations, offering the potential to be objective, fast, and reproducible. This work proposes a hybrid and fully automated methodology for bone age evaluation. The developed pipeline consists of three main stages: first, the images undergo a pre-processing phase, in which histogram equalization is applied to normalize brightness and contrast. Next, an object detection model (YOLO) is used to automatically locate and segment three bone regions of interest (joint, metacarpal, and carpal). Finally, a custom regression architecture is employed, in which three independent Convolutional Neural Networks (CNNs) process the images. The outputs of these networks are then fused through an adjusted weighted average and combined with the patient’s sex information, serving as input to a Multilayer Perceptron (MLP) that performs the final prediction. The final evaluation of the model, conducted on an isolated test set, resulted in a Mean Absolute Error (MAE) of 6.13 months and a Mean Absolute Percentage Error (MAPE) of 6.45%. These results validate the effectiveness of the proposed hybrid architecture, demonstrating that the integration of specialized regional features with demographic data yields a model with high generalization capability and competitive accuracy compared to the state of the art, making it clinically relevant for diagnostic support.