GalsenAI

Nous aurons le plaisir d’accueillir Dr Abdourahmane Baldé, docteur en informatique, spécialisé en intelligence artificielle appliquée à l’imagerie médicale et à la vision par ordinateur. Chercheur associé au laboratoire Cell Environment (France) et enseignant associé à l’Université numérique Cheikh Hamidou Kane (UN-CHK) au Sénégal, il coordonne depuis 2024 l’équipe sénégalaise de recherche en imagerie médicale du programme SPARK Academy. Ses travaux portent sur le développement de méthodes d’IA pour la détection et le diagnostic de la drépanocytose.

Il présentera l’article : SKDrepaData: A Multisite West African Blood Smear Dataset for Automated Sickle Cell Disease Detection and Classification, accepté au workshop MIRASOL 2026, dont voici l’abstract :

“Sickle cell disease (SCD) is one of the most prevalent inherited blood disorders worldwide, with the highest burden occurring in sub-Saharan Africa. Its diagnosis still relies heavily on manual microscopic examination of peripheral blood smears, a time-consuming and observer-dependent procedure, while publicly available microscopy datasets remain scarce and poorly representative of African clinical settings. We present SKDrepaData, a new dataset comprising 1,861 full-field blood smear microscopy images collected from more than 100 West African patients across three clinical sites using a standardized May-Grünwald-Giemsa (MGG) staining protocol. Red blood cells are annotated into three morphological categories (normal, sickle-shaped, and other abnormal cells), totaling 33,832 annotated cells with both bounding-box and cell-level labels. Annotation reliability was assessed through a five-expert cross-validation protocol, achieving a mean Intersection over Union (IoU) of 0.85 for object detection and a Fleiss’ kappa coefficient of 0.90 for morphological classification. To establish reproducible baseline benchmarks, we evaluate two previously published deep learning models for red blood cell detection and morphological classification on SKDrepaData. These models achieve a mAP@50 of 90.3% and a classification accuracy of 97.0%, respectively. The reported results provide reliable benchmark performances to facilitate future comparisons on this dataset. The dataset is available at https://doi.org/10.5281/zenodo.21365924.”