A Computationally efficient CNN-based Deep Learning Technique for Sickle Cell Detection
S Hemavarshini, Sathya Shree S, R. Arun · 2024
An amino acid change in the hemoglobin protein causes sickle cell disease (SCD), a prevalent hereditary illness that causes red blood cells to take on a sickle shape. These malformed cells cause severe health complications, including painful episodes, organ damage, increased infection risk, and a significantly reduced lifespan. Accurate and timely diagnosis is essential for timely therapy and better results for patients. This study investigates the potential of various lightweight deep learning models, MobileNetV1, MobileNetV2, MobileNetV3 Small, and EfficientNetV0 as an efficient and accurate underlying model for SCD detection. All of these trained them on the publicly available Sickle Cell Disease Dataset, where various augmentation methods are used to enhance the resilience and generalization of the models, such as rotating, shifting, shearing, zooming, brightness, and contrast adjustments. The MobileNetV1 has achieved better performance with an accuracy of 93.2% and F1Score of 95.43% with a lesser complexity with computational parameters of 4.855 Million and a model size of 18.52 MB. These findings highlight the potential of these lightweight models to revolutionize SCD screening and diagnosis in resource-constrained environments, offering faster, more accessible, and cost-effective diagnostic tools, ultimately improving healthcare outcomes for individuals with SCD.