Ensemble Deep Learning for Precise Thalassemia Classification in Blood Smear Images

Abstract

Thalassemia, particularly beta-thalassemia major, is a severe genetic blood disorder that presents significant challenges in diagnosis due to its similarity to other anemias. The lack of advanced diagnostic tools, especially in resource-limited regions, further complicates early detection. In this study, we present an ensemble deep learning approach for the precise classification of beta-thalassemia in blood smear images. After conducting multiple experiments, we selected three pre-trained convolutional neural networks—ResNet50, VGG16, and MobileNetV2—due to their complementary strengths in feature extraction and classification. These models were applied to a dataset consisting of 7,108 blood smear images, covering nine distinct erythrocyte classes. ResNet50 achieved a validation accuracy of 60.41% with a validation loss of 0.5956, but it struggled with overfitting. VGG16 outperformed ResNet50 with a validation accuracy of 65.89% and a validation loss of 0.3568, indicating more stable training. MobileNetV2, designed for efficiency in resourceconstrained environments, achieved a validation accuracy of 63.64% with a validation loss of 0.8648, though its loss curve suggested instability. The ensemble model demonstrated superior performance, achieving a validation accuracy of 75.61% and a validation loss of 0.5701, outperforming the individual models. Confusion matrix analysis and classification reports show that the ensemble model reduced misclassifications and improved accuracy across all thalassemia classes, particularly for difficult-to-classify cell types such as elliptocyte and teardrop. This ensemble system holds promise for assisting medical professionals in rural areas by enabling earlier diagnosis of thalassemia. Future work will focus on expanding the dataset and optimizing the model for deployment in real-world clinical settings, providing greater accessibility and improving diagnostic accuracy.