Object-detection annotations and multimodal dataset for the detection of microcytic hypochromic anemia
Description
This dataset supports the study "Towards Interpretable Deep Learning: An Explainable Multimodal Framework for the Detection of Microcytic Hypochromic Anemia." It comprises two components derived from the publicly available Anemic Red Blood Cell (AneRBC) benchmark dataset. The first is an object-detection annotation set of 94,221 labelled red-blood-cell instances across 1,002 whole-smear microscopy images, with each cell annotated as microcytic hypochromic, normal, or other using a documented three-class morphological protocol. The second is a multimodal, patient-level dataset for 417 patients (250 microcytic hypochromic anemic, 167 healthy controls), linking detection-derived morphometric image features with routine complete blood count (CBC) indices and the diagnostic label. Anemic patients were retained only where the morphological description confirmed a microcytic hypochromic picture; healthy controls were selected under strict criteria (normocytic normochromic morphology, hemoglobin ≥ 12 g/dL, MCV within 76–96 fL). The original smear images are sourced from the AneRBC benchmark and should be obtained from that source under its terms.
Files
Steps to reproduce
The dataset was constructed as follows. Peripheral blood smear images were obtained from the publicly available AneRBC benchmark dataset (Olympus DP27 camera, 40x magnification, Wright-Giemsa stain). Images were screened to remove blurred, under-stained, or heavily overlapping smears and standardized to 640x640 pixels with normalized intensities. For the object-detection dataset, 85 anemic patients' images were manually annotated on the Roboflow platform using a documented three-class protocol (microcytic hypochromic, normal, other), producing 94,221 labelled instances across 1,002 images; each cell was judged against fixed morphological criteria (reduced size and central pallor exceeding one third of the cell diameter for the microcytic class), with a medical doctor independently reviewing the labels. A YOLOv8 detector was then trained on these annotations (Ultralytics, AdamW optimizer, cosine schedule, 640x640 input, batch size 16) and applied by inference across all 417 patients. For each detected cell, morphometric descriptors (equivalent diameter, area, circularity, central pallor ratio, mean intensity) were computed by Otsu segmentation and aggregated per patient. These image-derived features were merged with the patients' CBC indices under strict clinical filtering (anemic: confirmed microcytic hypochromic morphology; healthy: normocytic normochromic morphology, hemoglobin >= 12 g/dL, MCV 76-96 fL) to form the multimodal dataset. Software: Python, Ultralytics YOLOv8, OpenCV, scikit-learn, SHAP. Full code is available in the associated repository.
Institutions
- National University of Sciences and TechnologyIslamabad, Islamabad