Hybrid CNN-Transformer Approach for Accurate Sickle Cell and Trait Detection in Blood Smears

Abstract

Sickle cell disease (SCD) is an inherited hemoglobinopathy that alters red blood cell morphology, causing anemia and vaso-occlusive complications. This study proposes a dual-backbone deep learning framework combining EfficientNetB2 or B4 and DenseNet169 to automate SCD detection from peripheral blood smear images. The architecture integrates Convolutional Block Attention Modules to refine spatial and channel features and transformer blocks to capture global dependencies through multi-head self-attention. The model is trained in two stages—frozen-backbone pretraining followed by fine-tuning—using Adam optimization and evaluated on the Digitized Thin Blood Films and ErythroSight datasets under binary and multi-class settings. Performance is assessed using accuracy, precision, recall, and ROC–AUC. Results demonstrate that the proposed hybrid model effectively combines local and global feature extraction, achieving superior performance to single-backbone networks and showing strong potential for scalable, automated SCD screening in low-resource settings.