[Paper Review] A Curated and Re-annotated Peripheral Blood Cell Dataset Integrating Four Public Resources

Lu Gan, Xi Li+1 more|arXiv (Cornell University)|Jul 18, 2024

Single-cell and spatial transcriptomicsBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper introduces TXL-PBC, a high-quality, re-annotated peripheral blood cell dataset integrating BCCD, BCD, PBC, and Raabin-WBC datasets after rigorous sample deletion, semi-automated labeling with YOLOv8n, and manual auditing. The resulting dataset—1008 training, 288 validation, and 144 test images—demonstrates superior label distribution and detection performance, with YOLOv8n achieving 0.964 mAP on the full set, significantly outperforming original datasets.

ABSTRACT

We present TXL-PBC, a curated and re-annotated peripheral blood cell dataset constructed by integrating four publicly available resources: Blood Cell Count and Detection (BCCD), Blood Cell Detection Dataset (BCDD), Peripheral Blood Cells (PBC), and Raabin White Blood Cell (Raabin-WBC). Through rigorous sample selection, semi-automatic annotation using the YOLOv8n model, and comprehensive manual review, we ensured high annotation accuracy and consistency. The final dataset contains 1,260 images and 18,143 bounding box annotations for three major blood cell types: white blood cells (WBC), red blood cells (RBC), and platelets. We provide detailed visual analyses of the data distribution, demonstrating the diversity and balance of the dataset. To further validate the quality and utility of TXL-PBC, we trained several mainstream object detection models, including YOLOv5s, YOLOv8s, YOLOv11s, SSD300, Faster R-CNN, and RetinaNet, and report their baseline performance. The TXL-PBC dataset is openly available on Figshare and GitHub, offering a valuable resource for the development and benchmarking of blood cell detection models and related machine learning research.

Motivation & Objective

• To address critical flaws in public blood cell datasets, including labeling errors, insufficient sample size, and poor data quality.
• To create a high-quality, sample-balanced dataset by integrating and re-annotating four public resources: BCCD, BCD, PBC, and Raabin-WBC.
• To improve the reliability and generalization of AI models for blood cell detection through enhanced data quality and consistency.
• To establish a benchmark dataset with standardized splits and comprehensive annotations for future research.
• To provide freely accessible, high-performance baseline models for detection tasks using YOLOv5 and YOLOv8 architectures.

Proposed method

• Conducted manual deletion of low-quality samples from BCCD and BCD datasets to improve data integrity.
• Performed semi-automatic labeling using YOLOv8n for all datasets, followed by manual audit to ensure annotation accuracy.
• Integrated the cleaned and re-labeled data from BCCD, BCD, PBC, and Raabin-WBC into a unified dataset with randomized shuffling and renaming for diversity.
• Split the final dataset into 1008 training, 288 validation, and 144 test images to ensure balanced evaluation.
• Trained multiple YOLO-based models (YOLOv5n/s/l, YOLOv8s/m) on TXL-PBC to establish performance baselines.
• Conducted comparative analysis of label distribution, bounding box area, and detection mAP across BCCD, BCD, and TXL-PBC to validate improvements.

Experimental results

Research questions

• RQ1 Does re-annotation and integration of public blood cell datasets significantly improve detection model performance?
• RQ2 How does the label distribution and bounding box area distribution in TXL-PBC compare to that of BCCD and BCD?
• RQ3 Can TXL-PBC serve as a reliable benchmark for training and evaluating new object detection models in blood cell analysis?
• RQ4 To what extent do different YOLO architectures perform on the TXL-PBC dataset compared to original datasets?
• RQ5 Can the integration of multiple datasets enhance model generalization and reduce labeling bias?

Key findings

• TXL-PBC exhibits a significantly wider distribution of label bounding box areas and a higher number of labels compared to BCCD and BCD, indicating improved data diversity and coverage.
• The YOLOv8n model trained on TXL-PBC achieved a mean average precision (mAP) of 0.964 on the full detection task, surpassing performance on both BCCD and BCD datasets.
• YOLOv8m achieved the highest mAP of 0.964 across all classes, with 0.960 for RBCs, 0.978 for WBCs, and 0.925 for platelets, demonstrating strong generalization.
• YOLOv5n, YOLOv5s, YOLOv5l, YOLOv8s, and YOLOv8m all achieved superior mAP scores on TXL-PBC compared to training on original BCCD and BCD datasets.
• The dataset's improved label consistency and reduced mislabeling significantly decreased model training errors and enhanced detection reliability.
• The TXL-PBC dataset is publicly available at https://github.com/lugan113/TXL-PBC_Dataset, enabling reproducible research and model benchmarking.