Anatomy-specific Progression Classification in Chest Radiographs via Weakly Supervised Learning.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Machine Learning for Healthcare

Background:

• Accurate classification of disease progression in chest radiographs is crucial for patient management.
• Current methods may lack efficiency in analyzing interval changes and detecting new pathologies.
• Weakly supervised learning offers a promising approach to leverage large datasets of radiology reports.

Purpose of the Study:

• To develop and evaluate a machine learning approach for classifying disease progression in chest radiographs.
• To utilize automatically derived weak labels from radiology reports for training.
• To enable monitoring of interval changes and detection of new pathologic conditions.

Main Methods:

• A retrospective study using a twin neural network architecture.
• Two-step weakly supervised learning: pretraining on MIMIC-CXR, fine-tuning on progression-labeled data.
• Evaluation using Area Under the Receiver Operating Characteristic (AUC) curves for classification and mean average precision for localization.

Main Results:

• The model achieved strong performance in classifying disease progression across six pathologic observations, with AUC scores ranging from 0.69 to 0.81.
• The algorithm demonstrated capability in localizing new progression areas, achieving mean average precision scores up to 0.34.
• Performance exceeded most baseline models in progression classification tasks.

Conclusions:

• Machine learning models can be effectively trained on large chest radiograph datasets using weakly supervised learning.
• These models can accurately classify disease progression and localize new findings.
• The developed approach facilitates monitoring interval changes and detecting new pathologic conditions in chest radiographs.