Processamento de imagens de radiografia de mãos para a detecção automática de maturidade óssea utilizando redes neurais convolucionais

Abstract

The evaluation of the bone age of pediatric patients is a standard clinical practice to determine their biological maturity, being an exam of great importance in identifying the conditions of growth and development in children and in predicting their future height. Typically, this process involves manual review of radiographic images of the hand using the Greulich-Pyle (GP) or Tanner-Whitehouse (TW) approach. The GP method employs a standardized hand atlas as the reference point for estimating bone age, while the TW approach uses a scoring system based on multiple regions of interest. However, this manual method is time consuming and prone to errors in skeletal age calculation. Therefore, an intelligent bone age assessment system using artificial intelligence could automate the repetitive task of image analysis. The proposed algorithm, developed in Python with the aid of its libraries, has two steps: first, image normalization and segmentation of the X-ray images, and then the use of a multi-input convolutional neural network regression architecture. This method has shown to be promising not only for theoretical demonstrations of academic content in the area of Digital Image Processing, but also for practical clinical use, with an average error rate of 7,31 months, making it possible to help medical professionals to assess bone age in a more accurate way.