Convolutional Neural Network Bone Age Assessment Radiograph
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated bone age assessment methods rely on hand-crafted features, leading to limited generalization and robustness, with existing systems rejecting radiographs due to noise and not utilizing discriminative carpal bones, and lack standardized datasets for fair comparison.
Innovation Solution
A system utilizing convolution neural networks to automatically generate bone age assessments from radiographs by processing images to enhance contrast, identifying hand and wrist regions, and determining bone age without user intervention, with trained models fine-tuned using normalized and annotated medical datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual bone age assessment using Greulich and Pyle atlas is used, then diagnostic accuracy is maintained, but time consumption and interrater variability increase significantly
Solution Approach 1:
The patent replaces the manual mechanical process of visual comparison with a trained convolutional neural network that automatically processes radiographs. The system substitutes human expert interpretation with an automated deep learning model that has been trained on normalized and annotated medical datasets, thereby eliminating time consumption and interrater variability while maintaining diagnostic accuracy.
Solution Approach 2:
The system enables self-service automation where the bone age assessment is performed autonomously without requiring human expert intervention. The trained model independently evaluates radiographs, generates bone age determinations, and produces reports, freeing clinicians from time-consuming manual assessments while preserving measurement precision.
2Extent of automation
If prior automated methods using hand-crafted features are used, then automation is achieved, but robustness decreases and systems reject radiographs due to noise
Solution Approach 1:
The patent replaces hand-crafted feature extraction with a trained convolutional neural network that automatically learns robust features from data. This deep learning approach substitutes manual feature engineering with automated feature learning, enabling the system to handle noisy images effectively without rejecting them, thereby maintaining high automation levels while improving reliability.
Solution Approach 2:
The system performs preliminary training on normalized and annotated medical datasets before deployment. This pre-training phase allows the model to learn from diverse, high-quality data, making it more robust to variations and noise in actual clinical images. The preliminary action of training on standardized datasets prepares the system to handle real-world variability without rejection.
3Productivity
If BoneXpert system is used, then automated assessment is provided, but carpal bones are not utilized and excessive noise causes rejections
Solution Approach 1:
The patent replaces the feature selection limitations of prior systems with a trained convolutional neural network that automatically identifies and utilizes discriminative carpal bones. The deep learning model substitutes manual feature selection with automated feature discovery, enabling effective utilization of carpal bones for young children while maintaining high productivity through automated processing.
Solution Approach 2:
The system changes the approach from fixed feature sets to learned features by training on normalized and annotated datasets. This parameter change allows the model to adaptively select and utilize discriminative features including carpal bones when appropriate, improving versatility without sacrificing assessment throughput. The training process optimizes feature utilization based on the specific characteristics of the input images.
Data Source
AI summary
In accordance with some embodiments, systems, methods and media for generating a bone age assessment. In some embodiments, a method comprises: receiving an x-ray image of a subject's left hand and wrist; converting the image to a predetermined size; identifying, without user intervention, a first portion of the image corresponding to the hand and wrist; processing the first portion of the image to increase contrast between bones and non-bones to generate a processed image; causing a trained convolution neural network to determine a bone age based on the processed image; receiving an indication of the bone age; causing the bone age to be presented to a user as the result of a bone age assessment; and causing the bone age and the image to be stored in an electronic medical record associated with the subject.