A Deep Learning-Based Feature Fusion Framework for Automated Thalassemia

Thalassemia is a blood disorder which restricts the body from producing enough hemoglobin, the protein containing iron. People with thalassemia must get frequent blood transfusions to maintain the hemoglobin levels in. In a thalassemic patient the blood smear images show microcytosis, hypochromia, anisocytosis, poikilocytosis, and codocytes due to defective hemoglobin synthesis. Early detection helps in the rapid and accurate identification of abnormal red blood cell morphology. Traditional analysis is prone to misinterpretation and time-intensive which leads to delayed treatment. In this approach a hybrid framework is built for automated thalassemia detection using deep learning approach. The EfficientNet algorithm extracts the local features and Vision Transformer extracts the global features. The combined features are utilized for binary classification of Thalassemia and Non-Thalassemia. This approach has provided an accuracy of 95%, 94% sensitivity, 96% specificity, 94% positive predicted value, 94% recall. This approach effectively automates thalassemia detection from blood smear images with balanced performance metrics and high accuracy.