Enhancing Clinical Trust in Deep Learning Models Through Grad-CAM-Based Explainability in Radiology Image Classification

Enhancing Clinical Trust in Deep Learning Models Through Grad-CAM-Based Explainability in Radiology Image Classification

Ali Alizadeh1

IOI Code : ABGZ-ABGABF
Publication : The third international conference on new approaches in engineering and basic sciences(icnabs.ir)
Abstract :
    Due to the opacity of deep learning decisions, deep learning models for image classification tasks in radiology have proven successful yet not implemented in clinics. According to a leading radiology artificial intelligence company, the problem of radiology artificial intelligence is not algorithm performance but the black box issue. The paper describes the use of Gradient-weighted Class Activation Mapping (Grad-CAM) as an explanatory mechanism to increase clinical trust in convolutional neural networks (CNN) predictions for chest X-rays. A framework is offered in which ResNet-50 and DenseNet-121 models are compared not only on conventional classification metrics, but also on the radiological relevance of their Grad-CAM heatmaps in relation to radiologist-identified regions of interest. In addition, our study takes note of the restrictions in the resolution and the scatter of attention in Grad-CAM. To improve the fidelity of localization, we also propose the stratification of aggregation across several levels. Furthermore, the implications of the European Union Artificial Intelligence Act (2024) and the General Data Protection Regulation on explainable AI in healthcare are considered in regulatory terms. The results imply that adding Grad-CAM visualizations to clinical decision-support systems could allow artificial intelligence to communicate more effectively with radiologists, as long as explainability is assessed with clinician-oriented metrics rather than computer-oriented metrics.

Keywords : explainable artificial intelligence grad cam deep learning radiology clinical trust convolutional neural networks