Consistent explainable image quality assessment for medical imaging

Medical image quality assessment is crucial, as poor-quality images can lead to misdiagnosis. Manual quality labeling is tedious for large studies and may produce misleading results. While automated analysis of image quality has been studied, little focus has been given to explaining and quantifying methodologies. This study proposes an explainable image quality assessment system, validated in two contexts: foreign object detection in Chest X-Rays (Object-CXR) and Left Ventricular Outflow Tract (LVOT) detection in Cardiac MRI.

Our explainable pipeline employs NormGrad, an algorithm that efficiently localizes image quality issues by comparing the classifier's saliency maps against several baseline saliency detectors. Additionally, a novel metric, the Difference of Means (DoM), is introduced to assess the consistency of saliency detectors across different network architectures.

We compare NormGrad with a range of saliency detection methods and demonstrate its superior performance in measuring the faithfulness of the saliency detectors. Specifically, NormGrad achieved a repeated Pointing Game score of 0.863 for Object-CXR and 0.778 for LVOT datasets, significantly outperforming other saliency detectors. Furthermore, our explainable pipeline shows strong consistency, with DoM scores of 0.001 for Object-CXR and 0.005 for LVOT datasets, indicating high reliability across different reproduced models. The code and experiments are publicly available at https://github.com/canerozer/explainable-iqa.

The proposed system, powered by NormGrad, significantly improves the reliability of automated medical image quality evaluations. The introduction of the Difference of Means metric offers a unique way to assess saliency detector consistency, supporting NormGrad's potential for widespread clinical adoption.