A Deep Framework for Bone Age Assessment based on Finger Joint Localization

Wu MinarXiv 2019 · arXiv

TL;DR

This paper introduces a novel deep learning framework for bone age assessment by focusing on the localization and analysis of finger joints. By extracting these discriminative regions, the authors achieve improved accuracy compared to traditional whole-hand approaches and even outperform other deep learning methods.

What Problem Does This Paper Address?

Bone age assessment is crucial for evaluating a child's skeletal maturity and diagnosing growth disorders. Traditional methods, like the Tanner-Whitehouse (TW) and Greulich and Pyle (GP) systems, rely on manual comparison of X-ray images with reference standards. However, these methods suffer from significant inter- and intra-observer variability, leading to inconsistent diagnoses. While deep learning has shown promise, many existing approaches treat the network as a black box, failing to incorporate domain knowledge from medical experts. This paper aims to bridge this gap by developing an automatic system that mimics expert observation by focusing on specific, informative regions – the finger joints.

Methods

The proposed framework consists of two main components: a joint localization strategy and a regression model.

  1. Joint Localization:
    • Bone Segmentation: A U-Net architecture is employed to segment the hand bones from the X-ray images. This is followed by a dense Conditional Random Field (CRF) to improve spatial coherence and segmentation quality.
  2. Finger Joint Extraction: After bone segmentation, the system identifies and localizes finger joints. This involves drawing rotated rectangles around distinct bone parts, computing their centroids, and then using these centroids to define the locations of the finger joints. Finally, fixed-size rotated rectangles around each joint are cropped to form the input for the regression model.
  3. Regression Model:
    • The cropped finger joint images are fed into a deep convolutional neural network (CNN).
  4. Gender information is incorporated by projecting it into a 32-dimensional space using a dense layer and concatenating it with the network's features.
  5. The network includes dense layers with dropout for regularization, and the final output is a predicted bone age value.
  6. To further improve robustness, the system also incorporates a separate network that processes the entire hand image. The final bone age prediction is obtained by averaging the outputs of the finger-joint-focused network and the whole-image network.

Datasets

  • Training and Validation: The Kaggle competition dataset held by the Radiological Society of North America (RSNA) is used. This dataset contains 12,611 training images, with 6,833 male and 5,778 female images. The authors used 10,089 images for training and the remainder for validation.
  • Testing: 200 testing images provided by RSNA are used for evaluation.

Key Insights and Ideas

  • Mimicking Expert Focus: The core idea is to leverage domain knowledge by identifying that finger joints are discriminative regions for bone age assessment, similar to how medical specialists focus on specific areas.
  • Information Fusion: Combining predictions from a model focused on localized finger joints with a model that analyzes the entire hand image leads to superior performance, suggesting that each approach captures complementary information.
  • Gender as a Feature: Incorporating gender information significantly improves the accuracy of bone age prediction, highlighting its importance in skeletal development.
  • Simplicity and Performance: The proposed method achieves state-of-the-art results with a simpler architecture compared to some existing complex systems, demonstrating the power of targeted feature extraction.

Key Results

  • The proposed finger joint localization strategy yields better performance than a whole-hand approach.
  • When fusing the results from the finger joint and whole-hand models, performance is further boosted.
  • With gender information, the proposed method achieved a Mean Absolute Difference (MAD) of 6.05 on the test set.
  • The proposed method outperforms the state-of-the-art method in [11] (which uses 32 VGG networks) with a simpler architecture (using two VGG models), achieving a MAD of 6.04 compared to 6.80 for the baseline VGG19 and 6.09 for the GRP method.

Limitations and Future Work

  • Information Loss: The authors acknowledge that some information might be lost during the finger joint localization process, which is why they propose fusing results from both localized and whole-image approaches.
  • Segmentation Accuracy: The performance of the system can be affected by the accuracy of the initial bone segmentation.
  • Future Work: While not explicitly detailed, the success of this approach suggests future work could explore other informative regions or more sophisticated fusion techniques. The paper also mentions the potential for further improvements by addressing segmentation inaccuracies.

Summary updated

BibTeX
@misc{min2019deep,
  author = {Wu Min},
  title = {{A Deep Framework for Bone Age Assessment based on Finger Joint Localization}},
  year = {2019},
  eprint = {1905.13124},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.13124},
}
Figure