Abstract
Background Sickle cell disease (SCD) is a genetic disorder affecting hemoglobin, leading to blood flow blockage and symptoms like pain and organ failure. It poses a significant global health burden, especially in regions such as sub-Saharan Africa and India. Early diagnosis is vital, but it is often hindered by traditional methods that require specialized resources. Artificial Intelligence (AI) is emerging as a solution, enhancing diagnostic techniques through faster and more accurate identification of SCD in blood smears, ultimately improving patient outcomes. This study focuses on developing AI applications to improve the early detection of SCD. Methods This study involved 81 participants. Of these, eight cases with thalassemia, hemoglobin E disease (HbE), and hemoglobin D disease (HbD) diagnoses were excluded from the study. The remaining 73, comprising 13 negative and 60 with sickle cell anemia (SS), sickle cell trait (AS), and AS+thalassemia, were included in the study. Each participant's blood sample underwent complete blood count (CBC), peripheral smear, and hemoglobin electrophoresis tests, with 730 data points generated from captured images analyzed by trained pathologists. The diagnosis by hemoglobin electrophoresis served as the gold standard, categorizing SCD as 1 and normal cells as 0. AI algorithms, including GoogLeNet and ResNet models, were developed using Python (Python Software Foundation, Delaware, United States) in Google Colab (Google LLC, Mountain View, California, United States), with performance assessed using sensitivity, specificity, recall, and F1-score metrics. Results Demographic data from participants indicates that the majority were aged 18-30, with 42 (57.53%) male participants. Analysis of CBC parameters revealed significant differences in hemoglobin, mean corpuscular volume (MCV), and mean corpuscular hemoglobin (MCH) between normal and SCD patients. Of those tested with hemoglobin electrophoresis, 13 (16.05%) were negative, while 60 (74.08%) tested positive for SCD, excluding cases with thalassemia, HbE, and HbD for AI analysis. A confusion matrix was used to assess the classification model's performance, focusing on true positives and negatives, as well as errors. Performance metrics such as accuracy, precision, recall, sensitivity, specificity, and F1-score were reported for three AI models, with ResNet50 convolutional neural network achieving the highest performance, followed by GoogLeNet and ResNet18. Conclusion This study confirms the high accuracy of AI in identifying sickle cells in blood smears. Despite challenges in validation, infrastructure, and adoption, AI-assisted screening could reduce diagnostic delays and improve outcomes in regions heavily impacted by SCD.