Improve bone age assessment by learning from anatomical local regions

Dong Wang, Kexin Zhang, Jia Ding, Liwei Wang

TL;DR

This paper introduces ALA-Net, a novel deep learning model for automated bone age assessment that mimics the clinically proven Tanner-Whitehouse (TW2) method by focusing on anatomical local regions of the hand. ALA-Net achieves state-of-the-art performance by jointly learning to detect these regions and predict bone age, offering a more interpretable and reliable solution for clinical use.

What Problem Does This Paper Address?

Skeletal bone age assessment (BAA) is a crucial diagnostic tool for evaluating a child's biological and structural maturation, aiding in the identification of growth disorders and endocrine issues. Clinicians traditionally rely on methods like the Greulich and Pyle (G&P) or Tanner-Whitehouse (TW2) methods, which involve visually analyzing X-ray images of the hand. The TW2 method, in particular, is highly regarded for its accuracy because it breaks down the hand into specific anatomical regions of interest (ROIs) and analyzes them individually. However, these manual methods are time-consuming and prone to subjective interpretation. While deep learning has shown promise in automating BAA, many existing models treat the hand image as a whole, failing to leverage the localized anatomical information that makes TW2 so effective. This paper aims to bridge this gap by developing a deep learning model that explicitly learns from anatomical local regions, inspired by the TW2 methodology.

Methods

The core of this work is the Anatomical Local-Aware Network (ALA-Net). ALA-Net integrates several key components:

  • Anatomical Local Extraction Module: This module is designed to learn the hand's structure and extract local information from specific anatomical regions. It takes features from an early stage of a ResNet-50 backbone and uses a Region Proposal Network (RPN) to identify potential ROIs. The top 17 ROIs are then selected and processed using ROI Align to extract local features. These features are then fed into a shared ROI head for further transformation. This module allows the network to learn anatomical structures and perform local analysis simultaneously, acting as a form of multi-task learning.
  • Anatomical Patch Training Strategy: Inspired by the TW2 method's intensive local analysis, this strategy uses the detected anatomical ROIs as "patches" for additional training. During training, the model predicts bone age on these individual ROI features, acting as a regularization technique. This encourages the network to learn the growth characteristics of each bone part, enhancing overall performance and interpretability.
  • Ordinal Regression: To address potential convergence issues with standard regression losses in multi-task learning, the paper employs ordinal regression. This transforms the bone age prediction problem into a series of binary classification tasks, where the model predicts whether the true age is greater than a given threshold. This approach is more stable and intuitive for age estimation.

The overall framework combines global features from the ResNet-50 backbone, local features from the extraction module, and gender information to make the final bone age prediction via a multi-layer perceptron.

Datasets

The experiments were conducted on the RSNA Pediatric Bone Age Challenge dataset. This dataset comprises 12,611 images for training, 1,425 for validation, and 200 for testing. The paper also mentions using hand keypoint annotations provided by a previous work [2] for training the RPN.

Key Insights and Ideas

  • Mimicking Clinical Expertise: The paper successfully demonstrates that deep learning models can effectively learn from localized anatomical information, mirroring the success of the TW2 method and making the model more interpretable for clinicians.
  • Multi-Task Learning for Structure and Age: By jointly learning to detect anatomical ROIs and predict bone age, the model significantly improves both tasks. Understanding hand structure aids in more accurate age prediction.
  • Anatomical Patch Training as Regularization: Using detected ROIs as training patches provides a powerful regularization mechanism, forcing the network to focus on specific bone development, which is crucial for accurate BAA.
  • End-to-End Learning with Human Priors: The ALA-Net is an end-to-end model that incorporates human prior knowledge (anatomical regions), making it more reliable and interpretable than purely data-driven, black-box approaches.

Key Results

ALA-Net achieved a Mean Absolute Error (MAE) of 3.91 months on the RSNA dataset, setting a new state-of-the-art for single-model performance. Ablation studies showed that:

  • The baseline ResNet-50 model achieved an MAE of 4.92.
  • Adding the anatomical local extraction module improved the MAE to 4.53.
  • The anatomical patch training strategy alone (using annotated ROIs) improved the MAE to 4.49.
  • Combining both the local extraction module and patch training yielded the best result of 3.91 MAE. The ROI detection component of ALA-Net also showed superior performance compared to a vanilla RPN, achieving higher AP scores.

Limitations and Future Work

  • Interpretability vs. Black Box: While ALA-Net is more interpretable than purely global models, the paper acknowledges that further work could explore deeper integration of medical knowledge into model design.
  • Data Dependency: The model's performance is dependent on the quality and quantity of the training data, particularly the hand keypoint annotations used for training the RPN.
  • Future Directions: The authors encourage further exploration of how to integrate medical knowledge into medical image analysis more broadly. The success of ALA-Net suggests that similar region-aware approaches could be beneficial in other medical imaging tasks.

Summary updated

Abstract

Skeletal bone age assessment (BAA), as an essential imaging examination, aims at evaluating the biological and structural maturation of human bones. In the clinical practice, Tanner and Whitehouse (TW2) method is a widely-used method for radiologists to perform BAA. The TW2 method splits the hands into Region Of Interests (ROI) and analyzes each of the anatomical ROI separately to estimate the bone age. Because of considering the analysis of local information, the TW2 method shows accurate results in practice. Following the spirit of TW2, we propose a novel model called Anatomical Local-Aware Network (ALA-Net) for automatic bone age assessment. In ALA-Net, anatomical local extraction module is introduced to learn the hand structure and extract local information. Moreover, we design an anatomical patch training strategy to provide extra regularization during the training process. Our model can detect the anatomical ROIs and estimate bone age jointly in an end-to-end manner. The experimental results show that our ALA-Net achieves a new state-of-the-art single model performance of 3.91 mean absolute error (MAE) on the public available RSNA dataset. Since the design of our model is well consistent with the well recognized TW2 method, it is interpretable and reliable for clinical usage.
BibTeX
@inproceedings{wang2020improve,
  author = {Dong Wang and Kexin Zhang and Jia Ding and Liwei Wang},
  title = {{Improve bone age assessment by learning from anatomical local regions}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
}
Figure