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

Sep 21, 2026

Uploaded by Aurelia

Deep learning image classifiers often achieve high predictive performance while remaining opaque about how decisions are formed. Correct outputs may rely on irrelevant cues, shortcut associations, peripheral structures, or acquisition artifacts instead of task-relevant regions. ReMoDEx provides a dataset-scale, stepwise explainability pipeline combining multiple local explainers with a global module to summarize relevance maps into strategy clusters. Applied to a VGG16 COVID-19 chest X-ray classifier, it reveals recurring central-vs-border decision strategies, validated by masked-image tests, and supports scalable assessment beyond accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Abstract

1. Introduction

ReMoDEx: A Local-to-Global Relevance-Based Model Decision Explainability Framework for large-Scale Image Datasets Abhay Kumar Pathaka, Mrityunjay Chaubeyb, Manjari Guptaa,* a Department of Computer Science, Institute of Science, Banaras Hindu University, Varanasi, India b University of Petroleum and Energy Studies, Dehradun, India * Corresponding author: Manjari Gupta ([manjari@bhu.ac.in](manjari@bhu.ac.in) ) Abhay Kumar Pathak ([pathakabhay@bhu.ac.in](pathakabhay@bhu.ac.in) ) Mrityunjay Chaubey ( [mrityunjay.chaubey@ddn.upes.ac.in](mrityunjay.chaubey@ddn.upes.ac.in)) Abstract: Deep learning image classifiers achieve strong predictive performance yet remain opaque in how decisions are formed. A model may predict correctly while relying on irrelevant cues, shortcut associations, peripheral structures, or device level artifacts instead of task relevant regions. On large scale datasets this opacity is especially problematic, since inspecting heatmaps one sample at a time cannot scale to thousands of predictions. We propose Relevance Based Model Decision Explainability (ReMoDEx), a framework for systematic, dataset scale assessment of model decision behaviour in image classification. ReMoDEx defines a stepwise pipeline: model inference, target class selection, relevance map generation, heatmap standardisation, similarity based grouping of patterns, cluster level interpretation, and spatial relevance assessment. Local methods GradCAM++, Integrated Gradients, Occlusion Sensitivity, and Layerwise Relevance Propagation are each combined independently with a single global module that summarises an entire set of relevance maps into a few decision strategy clusters, replacing sample by sample inspection with an automatic, scalable summary. To demonstrate ReMoDEx, we applied it to a VGG16 based classifier distinguishing COVID-19, Normal, Lung Opacity, and Viral Pneumonia. The classifier showed stable performance (86.27% test accuracy, 0.9624 test AUC) . However, each explainer combined with the global module consistently produced two recurring strategies: central thoracic region decisions and border/corner sensitive decisions, indicating possible shortcut learning that conventional metrics could not reveal. Masked image validation confirmed that model confidence and predicted class changed when central or peripheral regions were occluded. ReMoDEx thus provides a scalable relevance based decision assessment framework and an essential complement to accuracy based evaluation. Keywords: Explainable artificial intelligence; Chest X-ray classification; COVID-19 radiography; Clever Hans effect; Shortcut learning; Relevance-based decision; Large-scale model explainability; Model interpretability. 1. Introduction Deep learning image classifiers learn hierarchical visual representations directly from image data, allowing them to perform classification without depending on manually designed feature descriptors (LeCun et al., 2015; Litjens et al., 2017) . This representation-learning ability has made deep neural networks highly effective for image classification, but the model output usually provides only the predicted class or a probability score. The predicted label does not directly reveal which image regions, visual structures, or spatial patterns contributed to the decision. Consequently, a model can appear accurate at the output level while its internal decision process remains unclear. This lack of transparency matters because image classifiers may learn visual correlations that are useful for separating classes in the training data but are not necessarily related to the intended object, pathology, or task-specific evidence. Such correlations may arise from background regions, image borders, acquisition settings, preprocessing differences, text markers, or other dataset-dependent visual patterns. In these situations the prediction may be correct while the visual basis of the decision is not meaningful for the intended task. Con

FAQs

Why do high accuracy image classifiers still require decision explainability?

Because predictions can be correct while relying on irrelevant cues or unintended visual correlations, leaving the internal decision basis unclear. Accuracy and AUC do not verify whether task-relevant evidence was used.

What is ReMoDEx designed to do?

ReMoDEx enables systematic, dataset-scale assessment of model decision behaviour in image classification. It replaces sample-by-sample heatmap inspection with an automatic, scalable summary of relevance maps into strategy clusters.

How did ReMoDEx validate that the model used central or peripheral regions?

Masked image validation showed that model confidence and predicted classes changed when central thoracic areas or peripheral/border regions were occluded. The global module also consistently produced two recurring decision strategies.

Sep 21, 2026

Uploaded by Aurelia