MABAL: a Novel Deep-Learning Architecture for Machine-Assisted Bone Age Labeling.

Area of Science:

• Pediatric Radiology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Bone age assessment (BAA) is crucial for evaluating skeletal maturity in children.
• Traditional BAA methods, like the Greulich and Pyle atlas, are time-consuming.
• Existing computer-assisted detection (CAD) tools have limitations, and deep learning applications for BAA are emerging but face challenges with architecture and dataset size.

Purpose of the Study:

• To demonstrate the benefits of a customized neural network for bone age assessment.
• To show that advanced neural network architectures can be trained from scratch in medical imaging.
• To achieve superior performance compared to existing algorithms for BAA.

Main Methods:

• A customized 14-hidden layer neural network incorporating residual connections, inception layers, and spatial transformer layers was designed.
• A large dataset of 10,289 hand radiographs was utilized, separated into four cohorts (young/old males/females).
• Data augmentation and linear regression with mean square error loss were employed; Mean Absolute Error (MAE) was the primary performance metric.

Main Results:

• The customized neural network achieved aggregate validation and test set MAE of 0.637 and 0.536, respectively.
• This represents the best published performance for deep learning-based bone age assessment to date.
• Testing error consistently outperformed validation error, suggesting robust performance even with noisy training data.

Conclusions:

• Customized, purpose-built neural networks offer improved performance for bone age assessment over pre-trained models.
• State-of-the-art techniques like residual and inception architectures enhance prediction accuracy when trained from scratch.
• The study validates the feasibility of training small, customized CNNs with advanced strategies for medical imaging tasks, yielding significant accuracy improvements.