Visualizing artificial intelligence predictions for clinical experts − Detection of acute rejection from cardiac magnetic resonance images of heart transplant patients
Year: 2026
Type: article
Abstract: Background Most current convolutional neural network (CNN)-based artificial intelligence (AI) models have a limited ability to visualize how the resulting prediction is formed. This overshadows the integration of AI models into the clinical routine. This study aimed to develop a clinician-oriented workflow for visualizing CNN predictions in cardiac magnetic resonance (CMR) mapping, using acute heart transplant rejection as a clinical example. Methods We developed a pipeline to visualize image regions contributing to CNN predictions in a clinically interpretable format. As a proof of concept, the workflow was applied to a CNN detecting acute rejection from CMR T1 and T2 maps of heart transplant patients. The model used T1 and T2 maps from basal, mid-ventricular, and apical regions of the left ventricle. The operation of the model was visualized by using class activation mapping (ScoreCAM and GradCAM), and normalized occlusion maps. These methods were applied to input CMR images. The outputs were masked to the myocardium, transformed into a color-coded 16-segment cardiac model, and quantitatively compared with segmental ground-truth findings. Results The pipeline enabled automated analysis of CNN visualization activations and their presentation in a clinically familiar 16-segment model. Substantial intersegment variability was observed. Conclusion We presented a clinician-oriented workflow for visualizing and evaluating CNN predictions in CMR mapping. Using heart transplant rejection as an example, the approach transformed post-hoc AI visualizations into a clinically familiar cardiac segment model. Independent clinical validation isrequiredbefore thevisualizationoutputs can be linked toclinical phenomenaor used in clinical decision-making.
Source: Physica Medica
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,780
OpenAlex ID: W7237336367
