SKDrepaData-v2

SKDrepaData-v2 is a biomedical image dataset composed of 1,489 microscopic images of blood smears collected from 20 patients diagnosed with sickle cell disease (SCD). Samples were obtained from the UMRED laboratory (University of Thiès, Senegal) and the Children’s Hospital of Diamniadio (Dakar, Senegal), stained using the Giemsa technique, and digitized at 100x magnification using an optical microscope. The images were preprocessed with the Noise2Void [link] denoising model to enhance visual quality and reduce background noise. Manual annotation was performed by hematology experts, categorizing each visible cell into three main classes: normal cells, sickle cells, and other abnormal morphologies. The dataset reflects the natural class imbalance commonly observed in hematology, with a higher prevalence of normal cells. To mitigate training bias, two complementary strategies were used: data augmentation (random cropping and resizing, rotations, horizontal and vertical flips, color jitter, standardized normalization) and class-weighted sampling to balance rare cell types during training batches. The dataset is split as follows: 70% for training (1,131 images), 20% for validation (268 images), and 10% for testing (90 images).