Adversarial CAM Guidance for Chest X-Ray Classification: Reducing Framing Sensitivity with Mask Supervision

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computer vision

Background:

• Deep learning classifiers for chest X-rays often rely on background cues, leading to unreliable diagnoses.
• This framing sensitivity causes models to fail under distribution shifts, reducing trust.

Purpose of the Study:

• To develop a bio-inspired adversarial attention alignment training process for chest X-ray diagnosis.
• To enhance classifier decision-making by focusing on relevant evidence rather than image framing.

Main Methods:

• Trained a classifier with image-level labels, then used Class Activation Mapping (CAM) to generate attention heatmaps.
• Employed an adversarial approach with a discriminator to align CAM heatmaps with ground truth masks.
• Updated the classifier using a joint objective to maintain classification performance and encourage evidence-centered attention.

Main Results:

• The proposed method improved lung-focused attention and model robustness against framing sensitivity.
• New evaluation metrics, augmentation inconsistency and framing sensitivity, were introduced.
• The approach requires masks only during training, with no additional inference inputs.

Conclusions:

• Bio-inspired adversarial attention alignment effectively reduces reliance on background cues in chest X-ray diagnosis.
• This method enhances the trustworthiness and robustness of deep learning models for medical imaging.
• The technique offers a promising direction for developing more reliable AI diagnostic tools.