Manuscript received June 7, 2026; revised June 27, 2026; accepted July 13, 2026; published October 19, 2026.

Abstract—Recent advances in artificial intelligence have improved automated cardiovascular image analysis, but clinical use remains constrained by small annotated datasets, heterogeneous imaging sources, and limited generalizability across acquisition protocols. This study evaluates a modality-agnostic transfer learning pipeline for binary classification of echocardiographic and cardiac magnetic resonance images under data-limited conditions. Images were standardized using resizing, normalization, noise reduction, and data augmentation. A Visual Geometry Group 16 (VGG16) network pretrained on ImageNet was adapted with global average pooling, a dense classification block, dropout regularization, and a sigmoid output layer. Stratified 5-fold cross-validation was performed on the internal dataset. External evaluation used the Automated Cardiac Diagnosis Challenge Segmentation (ACDC) for magnetic resonance imaging and Cardiac Acquisitions for Multi-structure Ultrasound (CAMUS) dataset for an ejection fraction-derived echocardiographic assessment. The pipeline achieved an internal F1-Score of approximately 0.88 and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) of approximately 0.91. External F1-Scores were approximately 0.83 on ACDC and 0.80 on CAMUS, reflecting domain shift. Gradient-weighted class activation mapping indicated attention to anatomically meaningful cardiac regions. The results support this reproducible proof-of-concept approach, while multicenter validation remains necessary before clinical application.

Keywords—cardiac image classification, heterogeneous medical imaging, modality-agnostic learning, echocardiography, cardiac Magnetic Resonance Imaging (MRI), transfer learning, external validation, Gradient-weighted Class Activation Mapping (Grad-CAM)
 

Cite: Maxat Kabdullin*, Azat Kabdullin, Lyazat Naizabayeva, Vladimír Siládi, and Assiya M. Issayeva, "Modality-agnostic Transfer Learning for Heterogeneous Cardiac Image Classification," Journal of Image and Graphics, Vol. 14, No. 5, pp. 868-878, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).