ExplainabilitySafetyAccountabilityeess.IV
Medical Image Denosing via Explainable AI Feature Preserving Loss
Guanfang Dong, Anup Basu
October 31, 2023
International Conference on Smart Multimedia
10 citations
Key Takeaways
-
Medical image denoising requires protecting diagnostic features, not just removing noise: The paper identifies a critical gap in traditional denoising algorithms—they remove noise but may inadvertently eliminate or distort the medical features doctors rely on for diagnosis, potentially leading to misdiagnosis and legal liability.
-
Explainable AI can be repurposed as a quality control mechanism: Rather than using XAI only to explain model decisions after the fact, this work leverages gradient-based XAI methods during the denoising process itself to ensure that clinically important features remain consistent before and after noise removal.
-
The method was validated across diverse medical imaging scenarios: The approach was tested on three different medical image datasets with 13 different types of synthesized noise and artifacts, demonstrating that the solution generalizes across various real-world medical imaging challenges rather than working for just one specific case.
-
Balancing multiple objectives improves trustworthiness: By designing a loss function that simultaneously optimizes for noise removal, feature preservation, and model explainability, the method addresses the responsible AI principle that medical AI systems should be both effective and interpretable to clinicians.
Abstract
Denoising algorithms play a crucial role in medical image processing and analysis. However, classical denoising algorithms often ignore explanatory and critical medical features preservation, which may lead to misdiagnosis and legal liabilities. In this work, we propose a new denoising method for medical images that not only efficiently removes various types of noise, but also preserves key medical features throughout the process. To achieve this goal, we utilize a gradient-based eXplainable Artificial Intelligence (XAI) approach to design a feature preserving loss function. Our feature preserving loss function is motivated by the characteristic that gradient-based XAI is sensitive to noise. Through backpropagation, medical image features before and after denoising can be kept consistent. We conducted extensive experiments on three available medical image datasets, including synthesized 13 different types of noise and artifacts. The experimental results demonstrate the superiority of our method in terms of denoising performance, model explainability, and generalization.