Using Machine Learning Methods Incorporating Individual Reader Annotations to Classify Paediatric Chest Radiographs...

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Epidemiological research methodology

Background:

• Chest radiograph (CXR) interpretation in epidemiological studies faces significant inter-reader and intra-reader variability.
• This variability complicates cross-study comparisons and hinders efforts to track disease burden and evaluate interventions.
• Standardizing CXR interpretation is crucial for reliable epidemiological data.

Purpose of the Study:

• To explore machine learning models for standardizing CXR interpretation across studies.
• To evaluate the utility of incorporating individual reader annotations in models trained on multi-reader CXR datasets.

Main Methods:

• Convolutional neural networks (CNNs) were employed for CXR classification.
• Models were trained using WHO's standardized methodology for pediatric CXRs from seven low- to middle-income countries.
• Comparison between models predicting aggregate classification versus those incorporating individual reader predictions.

Main Results:

• Incorporating individual reader annotations improved classification accuracy by 3.4% (61% vs 59%).
• Model performance was notably higher for children over 12 months of age (68% vs 58%).
• CXRs classification accuracy varied by country, ranging from 45% to 71%.

Conclusions:

• Machine learning models effectively annotate CXRs, mitigating reader variability in epidemiological studies.
• Integrating individual reader annotations enhances the performance of ML models trained on multi-annotator CXR data.
• This approach offers a pathway to more consistent and reliable epidemiological findings from CXR studies.