Early Detection of Sickle Cell Anemia and Iron Deficiency Anemia from CBC and Peripheral Blood Smear Images

Abstract

Sickle Cell Anemia (SCA) is an inherited genetic blood disorder caused by mutations in hemoglobin genes. In this disorder, red blood cells become sickled and less flexible, leading to blocked blood flow in small vesicles, pain, and organ damage. Iron Deficiency Anemia (IDA) is the most common nutritional blood disorder, affecting millions worldwide. It is caused by insufficient iron availability, which lowers hemoglobin synthesis and impairs oxygen delivery. Both disorders are prevalent in resource-limited settings, and early detection is essential to reduce complications, prevent long-term health consequences, and alleviate the healthcare burden.Conventional diagnostic methods for both conditions depend on biochemical assays, imaging, or specialized techniques that are costly, infrastructure-dependent, and inaccessible in many rural and underserved areas. As a result, diagnoses are frequently delayed, increasing risks for vulnerable populations and leaving many cases untreated. This study presents an advanced, AI-based, unified diagnostic system for the early and accurate detection of both SCA and IDA. The system integrates two data-driven approaches combining peripheral blood smear (PBS) images with Complete Blood Count (CBC) data to detect SCA, and only CBC data to detect IDA. Hematological patterns and red cell morphology are analyzed using machine learning (ML) and deep learning (DL) techniques to provide reliable and precise diagnostic outcomes. For sickle cell anemia, the system was trained on 364 CBC records and 691 labeled PBS. The Conventional Neural Network model achieved 98.44% accuracy on image data, while the XGBoost model reached 98.63% accu- racy on CBC. When combined using a rule-based multi-modal fusion app roach, the overall diagnostic accuracy was 98.5%. For iron deficiency anemia detection, a large dataset of 15,300 records was collected from a clinical cohort, preprocessed, and analyzed using supervised learning. Random Forest with an accuracy of 99.98% outperformed the other models. With ROC–AUC was 0.9994, confirming excellent discriminative ability. To enable immediate diagnostic support, a Gradio-based user interface was developed, allowing clinicians to upload input data, receive real-time results, and download structured PDF reports. SHAP values were used to guarantee model interpretability, emphasizing features such as hemoglobin, MCV, MCH, and RDW, thereby enhancing physician trust and aligning with clinical reasoning. This study demonstrates that integrating routine clinical CBC and PBS with artificial intelligence (AI) can improve the speed, accuracy, interpretability, and accessibility of hematological diagnostics, particularly in low-resource healthcare environments.