Rethinking Annotation Granularity for Overcoming Shortcuts in Deep Learning-based Radiograph Diagnosis: A Multicenter...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Computer-Aided Diagnosis

Background:

• Deep learning (DL) models often exhibit shortcut learning, relying on superficial patterns rather than true diagnostic indicators in medical imaging.
• Chest radiograph interpretation is complex, with potential for bias in automated diagnostic systems.
• Fine-grained annotations, such as lesion localization, may offer a solution to improve DL model robustness.

Purpose of the Study:

• To evaluate if fine-grained, lesion-level annotations can overcome shortcut learning in DL-based chest radiograph diagnosis.
• To compare the performance of DL models trained with radiograph-level versus lesion-level annotations.
• To assess the generalizability of these models on external datasets.

Main Methods:

• Developed two DL models: CheXNet (radiograph-level annotations) and CheXDet (lesion-level bounding box annotations).
• Utilized a large dataset of 34,501 chest radiographs with annotations for nine conditions.
• Evaluated internal and external classification and lesion localization performance using ROC curve analysis and compared with radiologist performance.

Main Results:

• CheXDet demonstrated superior external classification performance compared to CheXNet, notably for fractures on the NIH Google and PadChest datasets.
• CheXDet achieved significantly higher lesion detection performance across multiple datasets and abnormalities, such as pneumothorax.
• Both models performed comparably to radiologists when provided with sufficient training data.

Conclusions:

• Fine-grained annotations effectively mitigate shortcut learning in DL models for chest radiography.
• Lesion-level annotations enhance the ability of DL models to identify correct pathological patterns, improving diagnostic accuracy and generalizability.
• This approach holds promise for developing more reliable and robust computer-aided diagnosis systems.