African Tropical Disease Diagnostic Image Dataset

This is a synthetic dataset generated from high-quality expert-labelled seed data. All records are algorithmically derived — statistical distributions, inter-field correlations, and annotation characteristics faithfully replicate real-world patterns from the source data, while ensuring no real individual, organisation, or transaction can be identified or reconstructed.

The African Tropical Disease Diagnostic Image Dataset contains 120K+ clinical images collected at partner hospitals and community health clinics across Nigeria, Ghana, Kenya, and Tanzania. The dataset covers 8 disease targets: malaria (Giemsa-stained blood smear microscopy), trachoma (external eye photographs, WHO TF/TI grading), cutaneous leishmaniasis (skin lesion photographs), schistosomiasis (urine microscopy), tuberculosis (chest X-ray), typhoid (clinical photograph panels), sickle-cell crisis (peripheral blood smear), and healthy/negative controls for each modality. Images were captured using standardised protocols on optical microscopes with smartphone adaptors, digital fundus cameras, and portable X-ray units.

Each image is labelled by at least two clinicians — a specialist physician and a trained clinical officer — using a consensus adjudication protocol. Labels include: disease class, severity or parasitaemia grade (modality-dependent), image quality score (1–5), and a deployment-readiness flag indicating whether the image meets quality thresholds for model training. Bounding box annotations are available for 40 % of images; the remainder carry image-level class labels only.

The dataset is split into train / validation / test sets stratified by disease, site, and image quality. A separate low-quality evaluation set is provided to measure model robustness under field conditions. Image modalities are stored in separate sub-directories with modality-specific metadata files. All images are provided as JPEG (photographs, eye images) or 16-bit TIFF (microscopy, X-ray) with JSON annotation sidecars in COCO format.