An Interaction-based Convolutional Neural Network (ICNN) Towards Better Understanding of COVID-19 X-ray Images

Explainability doesn't require sacrificing accuracy: The ICNN model achieved 99.8% accuracy on COVID-19 X-ray classification while maintaining interpretability, demonstrating that explainable AI systems can perform as well as traditional "black box" deep learning models.

Traditional CNN approaches may rely on irrelevant features: The paper challenges the standard practice of using pre-trained filters from other datasets, suggesting that conventional CNNs may incorporate assumptions about image features that aren't necessarily relevant to the specific medical task at hand.

Model-free approaches can identify what matters: Rather than relying on millions of pre-trained parameters, the ICNN uses an "Influence Score" to directly identify which parts of X-ray images are actually important for diagnosis, potentially making the decision-making process more transparent to clinicians.

AI screening tools could support clinical workflows: The authors propose this technology could "efficiently screen COVID-19 patients before human diagnosis," suggesting explainable AI systems might serve as a practical triage tool—though clinical validation and integration with existing diagnostic protocols would be necessary before deployment.

Interpretability is crucial for medical AI adoption: By demonstrating that an explainable model can match state-of-the-art performance, this work supports the argument that healthcare AI systems should prioritize interpretability so clinicians can understand and trust the model's recommendations.