ReMoDEx: A Local-to-Global Relevance-Based Model Decision Explainability Framework for large-Scale Image Datasets

Deep learning models for image classification often achieve high accuracy while relying on spurious correlations—such as image borders, background artifacts, or device-specific markers—rather than the intended clinical or task-relevant features. This phenomenon, known as shortcut learning or the Clever Hans effect, is difficult to detect using standard performance metrics like accuracy or AUC. The authors propose ReMoDEx (Relevance-Based Model Decision Explainability), a framework designed to scale the analysis of model decision-making from individual samples to entire datasets.

ReMoDEx operates through a multi-step pipeline: it generates local relevance maps for a trained classifier, standardizes these maps, and uses a global module to cluster them into distinct decision strategies. By grouping thousands of individual explanations into a few representative patterns, the framework allows researchers to identify whether a model is consistently focusing on relevant anatomical regions or if it is frequently defaulting to non-relevant image artifacts.

The authors demonstrated ReMoDEx using a VGG16-based classifier trained on a multiclass chest X-ray dataset (COVID-19, Normal, Lung Opacity, and Viral Pneumonia). While the model achieved strong performance (86.27% accuracy), the ReMoDEx analysis revealed that the model consistently employed two distinct decision strategies: one focused on the central thoracic region and another focused on peripheral border and corner regions.

Through masked image validation, the researchers confirmed that the model's confidence and classification output were sensitive to these peripheral regions. This suggests that the model was exploiting non-clinical, shortcut-sensitive cues to achieve its high performance. The study highlights that conventional metrics are insufficient for detecting these heterogeneous decision strategies, making ReMoDEx a necessary tool for auditing model reliability in high-stakes domains like medical imaging.

As deep learning models are increasingly deployed in clinical settings, relying solely on accuracy metrics poses a significant risk. A model might perform well on internal test sets by 'cheating' via dataset-specific artifacts, only to fail when deployed in real-world environments with different acquisition protocols. ReMoDEx provides a systematic, scalable way to verify that a model's 'reasoning' aligns with human-interpretable, task-relevant evidence, thereby improving the trustworthiness and robustness of AI-assisted diagnostics.