Self-eXplainable AI for Medical Image Analysis: A Survey and New Outlooks

Built-in explanations are more trustworthy than after-the-fact ones: The paper argues that Self-eXplainable AI (S-XAI), which incorporates explainability during model training, is superior to post-hoc XAI techniques that try to explain black-box models after they're already trained. This matters because explanations generated after the fact may not accurately reflect how the model actually made its decisions.

Medical imaging needs AI systems you can understand and verify: In high-stakes medical applications, transparent AI isn't just nice to have—it's essential. S-XAI approaches help ensure that when an AI system recommends a diagnosis based on a medical image, doctors can understand the reasoning and verify it's sound, supporting both trustworthiness and accountability.

Multiple explanation approaches serve different purposes: The survey identifies three complementary ways S-XAI systems can explain themselves: by highlighting which features in an image matter most (input explainability), by showing which parts of the model's internal logic are active (model explainability through attention and concepts), and by providing text descriptions or showing what would need to change for a different diagnosis (output explainability).

The field is rapidly developing with over 200 documented approaches: The breadth of research reviewed (200+ papers across various medical imaging modalities and clinical applications) indicates this is an active and maturing area, suggesting that explainable medical AI systems are becoming increasingly practical and available.

Assessing explanation quality remains a challenge: The paper notes that while S-XAI methods are advancing, the field still needs better standards for determining whether an explanation is actually good, reliable, and useful for clinicians—a critical gap for responsible deployment in healthcare.