Simulation-Trained Deep Learning for Automated Cell-Based HLA Antibody Assay Interpretation in Pre-Transplant Diagnostics - Image Datasets

Dataset description This Zenodo record contains standardized HDF5 image datasets used for the development, training, validation, and benchmarking of automated cell-segmentation workflows for image-based diagnostic assays. The deposited files are reduced, shareable subsets of the corresponding full-length development datasets. For each HDF5 dataset, the first 100 image samples were retained. Subsetting was performed along the sample dimension; all tiles belonging to a selected image were preserved. Consequently, datasets containing one tile per image retain a shape of [100, 1, ...], whereas tiled datasets retain their complete tile dimension, for example [100, 20, ...]. Full datasets are available upon reasonable request. Dataset types The collection includes complementary datasets representing different stages and sources of image-analysis development: Simulated image datasets These datasets contain synthetically generated assay images with corresponding pixel-level ground truth. They were created using configurable scene, acquisition, and camera-style simulations to represent variation in cell morphology, density, illumination, image quality, and acquisition characteristics. Depending on the dataset configuration, each simulated scene is represented either by a single processed image or by multiple partially overlapping tiles. Background-image dataset This dataset contains background-only image tiles without target cells. It was generated to expose the segmentation model to representative non-cellular image structures, acquisition noise, illumination variation, and potential background artefacts. The corresponding segmentation targets and instance-label maps contain no foreground objects. External-image validation dataset This dataset contains real assay images acquired independently of the simulated training data. The images were divided into tiles while retaining the associated segmentation masks and image-level metadata. It was used to evaluate whether segmentation workflows trained primarily on simulated data generalize to images acquired under real experimental conditions and with different imaging systems. Human-annotated image dataset This dataset contains real assay images with human-derived reference annotations. These samples provide independent ground truth for comparing automated cell detection and segmentation with expert assessment.