Deep Learning–Based Stratification of Iron Overload in Thalassemia Patients

Iron overload caused by frequent blood transfusions is one of the most serious problems in patients with transfusion-dependent thalassemia, with the heart and liver being the most affected organs. Excess iron accumulation can lead to irreversible organ damage, including cardiomyopathy, arrhythmias, liver fibrosis, and ultimately organ failure, making early and accurate assessment vital for patient management. Magnetic resonance imaging (MRI), particularly T2* mapping, is the current gold standard for noninvasive quantification of tissue iron concentration. However, manual interpretation of T2* images is time-consuming, subject to interobserver variability, and dependent on radiologic expertise. This study proposes a deep learning-based framework utilizing the EfficientNet-B2 architecture for the automated classification of cardiac and hepatic iron overload severity. The model was trained and evaluated using the CHMMOTv1 dataset, which includes annotated T2* MRI scans from thalassemia patients with varying levels of iron deposition. The model achieved classification accuracies of 64% for cardiac iron overload and 65% for hepatic iron overload. The area under the curve (AUC) values were 0.83 for cardiac and 0.80 for hepatic predictions, indicating strong and moderate discriminatory power, respectively. These findings underscore the feasibility and clinical potential of AI-assisted MRI analysis in improving diagnostic accuracy, reducing manual workload, and supporting timely, objective risk stratification. The framework sets the stage for further development of robust AI tools that can enhance the quality and scalability of iron overload monitoring in routine thalassemia care, particularly in resource-limited settings.