Automated Bone Age Assessment: A New Three-Stage Assessment Method from Coarse to Fine

Xinzheng Xu, Huihui Xu, Zhongnian LiHealthcare 2022 · DOI

TL;DR

This paper introduces a novel three-stage automated system for Bone Age Assessment (BAA) using X-ray images. By combining object detection (YOLOv5) with fine-grained classification (Context-Aware Attention Pooling), the system accurately identifies and grades 13 key hand bones, leading to more precise bone age estimations than traditional methods.

What Problem Does This Paper Address?

Determining a child's biological age, or "bone age," from X-ray images is crucial for monitoring growth and diagnosing developmental issues. However, the current manual assessment by radiologists is time-consuming, subjective, and prone to inter-observer variability. While deep learning has shown promise in automating this, existing methods often lack interpretability, providing only a final bone age value without detailing the maturity of individual bones, which is important for clinical decision-making. This paper aims to bridge this gap by developing an accurate, interpretable, and automated BAA system.

Methods

The proposed solution is a hierarchical, three-stage approach:

  1. Object Detection (Coarse Stage): The system first employs YOLOv5, a state-of-the-art object detection model, to precisely locate and extract 13 Regions of Interest (ROIs) corresponding to specific bones in the left hand X-ray. This stage ensures that the subsequent analysis focuses on the relevant anatomical structures.
  2. Fine-Grained Classification (Fine Stage): For each of the 13 detected ROIs, a fine-grained classification model is used to determine its maturity grade. The authors utilize a Context-Aware Attention Pooling (CAP) module, which is designed to capture contextual information at the pixel level, enabling more accurate differentiation between subtle variations in bone maturity. This is a key innovation, as traditional classifiers struggle with these fine distinctions.
  3. Bone Age Estimation (Final Stage): Finally, the graded maturity of the 13 bones is fed into the established Tanner-Whitehouse (TW3) methodology to compute the overall bone age. This stage leverages a well-validated clinical standard for the final estimation.

This cascaded approach allows for a more robust and detailed assessment, moving from identifying the general areas (coarse) to classifying their specific developmental stages (fine).

Datasets

The study utilized two datasets:

  • Public Dataset: The 2017 Pediatric Bone Age Challenge dataset from the Radiological Society of North America (RSNA), containing 12.6k left hand X-ray images.
  • Clinical Dataset: A newly established dataset comprising 2518 left hand X-ray images from Xuzhou Central Hospital. This dataset was meticulously annotated by professional radiologists, providing detailed labels for 13 bone categories and their grades, resulting in 32,734 fine-grained classification samples after cropping ROIs.

Key Insights and Ideas

  • Hierarchical Approach is Key: Breaking down BAA into object detection and fine-grained classification stages, rather than a single end-to-end model, significantly improves accuracy and interpretability.
  • Context Matters for Fine-Grained Classification: The CAP module's ability to leverage contextual information is crucial for distinguishing subtle differences in bone maturity, outperforming simpler classification methods.
  • Interpretability Enhances Clinical Utility: Providing not just a bone age value but also the detection and maturity grade of individual bones makes the system more valuable to clinicians, allowing them to investigate specific developmental details.
  • YOLOv5 for ROI Extraction: Using YOLOv5 for ROI detection proved highly effective, achieving high accuracy and avoiding the need for complex foreground-background segmentation.

Key Results

  • High ROI Detection Accuracy: YOLOv5 achieved an impressive detection accuracy of 98.93% for the 13 ROIs.
  • Superior Bone Grade Classification: The proposed CAP-based fine-grained classification achieved an average accuracy of 86.93% on the clinical dataset, significantly outperforming other compared methods.
  • Competitive Bone Age Estimation:
    • On the clinical dataset, the Mean Absolute Error (MAE) for bone age estimation was 7.68 months.
  • On the public RSNA dataset, the MAE was 6.53 months.
  • Comparison to Existing Methods: The paper shows that their three-stage method is superior to existing literature methods, particularly in accuracy, and is more accurate than traditional manual assessment.

Limitations and Future Work

  • Uneven Sample Distribution: The clinical dataset exhibited an uneven distribution of samples, with fewer images for younger children (0-5 years old). This imbalance is acknowledged as a potential factor affecting accuracy, especially for younger age groups.
  • Future Work: The authors plan to develop a more suitable framework to address the issue of imbalanced sample distribution, aiming to further improve the accuracy of bone age assessment, particularly for younger children.

Summary updated

Abstract

Bone age assessment (BAA) based on X-ray imaging of the left hand and wrist can accurately reflect the degree of the body's physiological development and physical condition. However, the traditional manual evaluation method relies too much on inefficient specialist labor. In this paper, to propose automatic BAA, we introduce a hierarchical convolutional neural network to detect the regions of interest (ROI) and classify the bone grade. Firstly, we establish a dataset of children's BAA containing 2518 left hand X-rays. Then, we use the fine-grained classification to obtain the grade of the region of interest via object detection. Specifically, fine-grained classifiers are based on context-aware attention pooling (CAP). Finally, we perform the model assessment of bone age using the third version of the Tanner-Whitehouse (TW3) methodology. The end-to-end BAA system provides bone age values, the detection results of 13 ROIs, and the bone maturity of the ROIs, which are convenient for doctors to obtain information for operation. Experimental results on the public dataset and clinical dataset show that the performance of the proposed method is competitive. The accuracy of bone grading is 86.93%, and the mean absolute error (MAE) of bone age is 7.68 months on the clinical dataset. On public dataset, the MAE is 6.53 months. The proposed method achieves good performance in bone age assessment and is superior to existing fine-grained image classification methods.
BibTeX
@inproceedings{xu2022automated,
  author = {Xinzheng Xu and Huihui Xu and Zhongnian Li},
  title = {{Automated Bone Age Assessment: A New Three-Stage Assessment Method from Coarse to Fine}},
  booktitle = {Healthcare},
  year = {2022},
  doi = {10.3390/healthcare10112170},
}
Figure