Bone Age Assessment (BAA) using Convolutional Transformer and MultiLayer Perceptron

Bone Age Assessment (BAA) is an important tool doctors use in the medical field to determine the growth and development of bones in a child. The Bone Age Assessment Report is mandatory for pediatricians and endocrinologists to address the child’s health. The most commonly used BAA technique is hand & wrist radiography, a technique established around 60 years ago. The traditional methods of the BAA were fully dependent on expert clinicians to identify the result manually. These methods were time-consuming and also produced inaccurate results. For BAA using hand and wrist radiography images, we provide a unique method in this paper that uses a convolutional transformer architecture. Our model consists of an embedding block and transformer encoder blocks, integrating both spatial relationships and global dependencies crucial for accurate estimation. Preprocessing techniques including image resizing, histogram equalization, noise removal, and adaptive masking are employed to enhance image quality and feature extraction. The convolutional token embedding block divides the image into fragments, preserving spatial connections, while the transformer encoder block incorporates self-attention mechanisms to capture global dependencies among bone regions. Multi-headed self-attention enables the model to focus on complex patterns within X-ray images, enhancing accuracy in bone age estimation. The output of the transformer blocks is refined through a multilayer perceptron, incorporating gender information for improved prediction. Validation techniques ensure the model’s robustness and prevent overfitting. Experimental results demonstrate the effectiveness of our proposed convolutional transformer model in accurately predicting bone age, offering potential applications in clinical practice for pediatric growth assessment.