Morphological Feature-Based CNN Model for the Sickle Cell Disease Diagnosis from Blood Smear Images

Abstract

Sickle Cell Disease (SCD) diagnosis traditionally relies on manual blood smear evaluation, which is time-consuming and subject to observer variability. In this study, we developed and evaluated a convolutional neural network (CNN) model aimed at classifying erythrocyte morphologies from microscopic blood smear images. We focused on differentiating various important red blood cell types, including sickle-shaped, oval, and discocyte forms, using a carefully selected subset of annotated images from the publicly accessible ErythrocytesIDB dataset. To separate and segment individual cells, preprocessing techniques included edge detection, morphological operations, and thresholding. The model was trained and tuned with a combination of texture-based, color-based, and geometric features. The proposed CNN exhibited competitive performance in comparison with common architectures, including ResNet50, VGG19, and DenseNet121, with impressive accuracy, precision, and recall values. While preliminary, these results suggest that machine learning models could support automated and scalable SCD screening, especially in resource-constrained clinical settings.