AI Safety, Alignment & Interpretability
Pixel-level interpretability model outperforms Grad-CAM on every tested metric in a COVID-19 chest X-ray study
In a 2025 paper published in Machine Learning and Knowledge Extraction, a peer-reviewed journal, researchers at Université du Québec à Chicoutimi reported that a new pixel-level interpretability model called PLI outperformed Grad-CAM on every measured dimension. The comparison used more than 1,000 labeled COVID-19 chest X-ray images from three public datasets. PLI produced more detailed heatmaps that highlighted localized diagnostic features. It also recorded higher structural similarity, lower error, and faster prediction times than Grad-CAM.
What did the study compare?
PLI stands for Pixel-Level Interpretability. It is a model that adds fuzzy logic, a form of reasoning that handles degrees of uncertainty, to its visual and numerical explanations. Grad-CAM, short for Gradient-Weighted Class Activation Mapping, is a common method that shows the broad regions of an image that influenced a model's decision. The experiments used the VGG19 convolutional neural network, a standard image-recognition model.
What did the paper find?
The paper reports that PLI significantly outperformed Grad-CAM in every measured dimension. PLI produced detailed pixel-level heatmaps. It scored higher on structural similarity (SSIM), a measure of how close two images are. It had lower mean squared error (MSE), a measure of average pixel error. It also ran faster during inference, the step where a trained model makes a prediction from an image. The researchers write that Grad-CAM's explanations often lacked the detail needed for clinical reliability. They say PLI's fuzzy logic can deliver explanations that align with clinical expectations.
Why does that matter in medical imaging?
A model that classifies a chest X-ray does not automatically show where it looked. Interpretability is how well a person can understand the model's reasoning. The paper argues that fine-grained heatmaps could help a clinician check whether the model focused on a relevant part of the chest. If the explanation is too coarse, that check becomes harder.
FAQ
What is Grad-CAM?
Grad-CAM is a common method for showing which parts of an image influenced a model's classification. This paper reports that its explanations were often less detailed than PLI's.
Which data did the study use?
The study used more than 1,000 labeled COVID-19 chest X-ray images from three publicly available datasets.
What does interpretability mean here?
Interpretability means how well a person can see and understand why an AI model made a decision. In this study, it was compared through visual heatmaps, similarity scores, error scores, and diagnostic precision.
Limits and what's next
The findings come from one comparative study of COVID-19 chest radiographs. The abstract describes a computational comparison on public chest X-ray datasets. The data are more than 1,000 labeled COVID-19 chest X-ray images from three public datasets. The design uses the VGG19 image model to compare PLI and Grad-CAM on interpretability, structural similarity, diagnostic precision, mean squared error, and computational efficiency. The abstract reports no independent replication and no test with practicing clinicians. The abstract does not state whether PLI's alignment with clinical expectations was externally validated. The abstract suggests possible broader use in other high-stakes areas, but it does not report tests beyond the COVID-19 chest X-ray datasets.
Source: Mohammad Ennab and Hamid Mcheick, “Advancing AI Interpretability in Medical Imaging: A Comparative Analysis of Pixel-Level Interpretability and Grad-CAM Models,” Machine Learning and Knowledge Extraction, 2025.