Blood_5_for_classification

# Datasets ## 1. Classification WBC — Blood5 (Self-Collected) ### Overview Blood5 is a **self-collected peripheral white blood cell (WBC) microscopy dataset** for five-class classification. It covers all five major mature leukocyte subtypes seen in clinical hematology: basophil, eosinophil, lymphocyte, monocyte, and neutrophil. ### Scale | Split | Samples ||-------|--------:|| Train | 20,697 || Test | 5,175 || **Total** | **25,872** | ### Class Distribution | Index | Class | Train | Proportion ||:-----:|-------|------:|-----------:|| 0 | Basophil | 424 | 2.0% || 1 | Eosinophil | 472 | 2.3% || 2 | Lymphocyte | 3,969 | 19.2% || 3 | Monocyte | 2,054 | 9.9% || 4 | Neutrophil | 9,638 | 46.6% | > The dataset exhibits **severe natural class imbalance** — neutrophils dominate while basophils and eosinophils are rare (<5% combined) — faithfully reflecting real-world clinical prevalence. ### Collection & Preprocessing 1. Blood smears prepared and stained following standard hematology protocols.2. Images captured via optical microscope with digital camera.3. Converted to RGB and resized to **150×150** pixels.4. Stored in CIFAR-10 binary pickle format with 80:20 stratified split (random seed = 42).5. Online augmentations: `RandomResizedCrop` (224×224), `RandAugment`, `Cutout`. ### File Format | File | Content ||------|---------|| `batches.meta` | Label name mapping || `data_batch_1` | Train batch 1 (10,000) || `data_batch_2` | Train batch 2 (10,000) || `data_batch_3` | Train batch 3 (697) || `test_batch` | Test set (5,175) | Each image: flattened 1D array (150×150×3 = 67,500 elements) → reshaped to H×W×C on load. ### Subset: Blood3 A three-class subset (lymphocyte, monocyte, neutrophil) is also provided for fast prototyping. --- ## 2. Segmentation WBC (Public) ### Overview Two publicly available WBC **segmentation** datasets from Zheng et al. (2018), used for evaluating cell segmentation methods. These datasets differ substantially in color, cell morphology, and background — offering a robust test of generalization. ### Ground Truth Masks are manually annotated by domain experts with three regions: | Intensity | Region ||:---------:|--------|| White | Nucleus || Gray | Cytoplasm || Black | Background (incl. red blood cells) | ### Dataset 1 | Property | Value ||----------|-------|| Source | Jiangxi Tecom Science Corp., China || Samples | 300 || Resolution | 120×120 || Color depth | 24-bit || Microscope | Motic Moticam Pro 252A + N800-D autofocus || Staining | Novel rapid hematology reagent || Appearance | Yellowish background | ### Dataset 2 | Property | Value ||----------|-------|| Source | CellaVision blog || Samples | 100 || Resolution | 300×300 || Appearance | Purplish cells, many surrounding red blood cells | ### Class Labels (Auxiliary) Labels 1–5 correspond to neutrophil, lymphocyte, monocyte, eosinophil, and basophil. **However**, these labels were added by the authors and not verified by domain experts — use with caution for classification purposes. ### Citation ```bibtex@article{Zheng2018, title = {Fast and Robust Segmentation of White Blood Cell Images by Self-supervised Learning}, author = {Zheng, Xin and Wang, Yong and Wang, Guoyou and Liu, Jianguo}, journal = {Micron}, volume = {107}, pages = {55--71}, year = {2018}, publisher = {Elsevier}, doi = {10.1016/j.micron.2018.01.010}}``` > **Acknowledgment**: Dataset 1 → Jiangxi Tecom Science Corporation; Dataset 2 → CellaVision blog.