A SMOTE-Based Deep Learning Approach for Sickle Cell Detection in Low-Resolution, Class-Imbalanced Microscopic Images (SDL-SCD)

sickle cell disease is a genetic condition characterized by abnormal red blood cell morphologies. It can be quite challenging to identify and monitor its response to treatment. Although deep learning-based models exhibit great potential in medical image processing, existing approaches often fail to cope with variability in sickle cell morphology. Additionally, publicly available sickle cell datasets tend to have a few samples with imbalanced classes. To mitigate the above challenges, we propose using the synthetic minority sampling technique (SMOTE) mechanism to handle class imbalances and a deep CNN architecture that aims to capture complex patterns and descriptive features in a newly created low-resolution sickle cell dataset from hospitals in eastern Uganda. This could help improve the efficiency of the diagnosis and classification of the disease. We performed experiments and examined several algorithms in the literature for related tasks. Based on the evaluation results, the proposed SMOTE-based DL-SCD outperforms the best baseline, its variant without the SMOTE component, with a 2.06% increase in classification accuracy. SDL-SCD could help to conveniently and early detect sickle cell anemia, especially in low-developed settings where medical services are constrained. Our code is accessible at github.com.