Related Experiment Video
Updated: Oct 14, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
7.0K
Deep Learning Algorithms for Interpretation of Upper Extremity Radiographs: Laterality and Technologist Initial
Paul H Yi1, Patrick S Malone2, Cheng Ting Lin3
1Department of Radiology and Nuclear Medicine, University of Maryland Medical Intelligent Imaging Center (UM2ii), University of Maryland School of Medicine, 670 W Baltimore St, First Fl, Rm 1172, Baltimore, MD 21201.
AJR. American Journal of Roentgenology
|November 10, 2021
Summary
Convolutional neural networks (CNNs) for identifying upper extremity radiographic abnormalities performed well. Covering radiograph labels improved CNN performance and shifted focus to bone structures, highlighting potential data confounding factors.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging Analysis
Background:
- Convolutional neural networks (CNNs) show promise in analyzing medical images.
- Previous studies indicate CNNs may rely on non-pathological image features, such as labels.
- Understanding these biases is crucial for developing reliable AI diagnostic tools.
Purpose of the Study:
- To evaluate the impact of radiograph labels on CNN performance for upper extremity abnormality detection.
- To investigate whether obscuring labels redirects CNN attention to relevant anatomical structures.
- To assess the diagnostic value of radiograph labels themselves.
Main Methods:
- Training CNNs to detect abnormalities on upper extremity radiographs.
- Comparing CNN performance with and without radiograph labels visible.
- Analyzing CNN attention using label-only images.
Main Results:
- CNNs achieved an Area Under the Curve (AUC) of 0.844 when labels were visible.
- Obscuring labels improved AUC to 0.857 (p = .02) and shifted CNN focus from labels to bones.
- CNNs achieved an AUC of 0.638 using only radiograph label images.
Conclusions:
- Radiograph labels can influence CNN decision-making in abnormality detection.
- Obscuring labels can improve CNN accuracy and direct attention to clinically relevant features.
- Careful data curation is necessary to mitigate confounding factors in AI development for radiology.
