[Paper Review] WBCAtt: A White Blood Cell Dataset Annotated with Detailed Morphological Attributes

Satoshi Tsutsui, Winnie Pang+1 more|arXiv (Cornell University)|Jun 23, 2023

Digital Imaging for Blood Diseases4 citations

TL;DR

This paper introduces WBCAtt, the first public dataset of 10,298 white blood cell images annotated with 11 detailed morphological attributes—such as nucleus shape, cytoplasm texture, granule type, and chromatin density—derived from pathologist consultations and medical literature. The dataset enables explainable AI in hematology by supporting high-accuracy attribute prediction (e.g., 99.96% F1 for granule color 'red' using ViT-Base), advancing interpretability in WBC classification models.

ABSTRACT

The examination of blood samples at a microscopic level plays a fundamental role in clinical diagnostics, influencing a wide range of medical conditions. For instance, an in-depth study of White Blood Cells (WBCs), a crucial component of our blood, is essential for diagnosing blood-related diseases such as leukemia and anemia. While multiple datasets containing WBC images have been proposed, they mostly focus on cell categorization, often lacking the necessary morphological details to explain such categorizations, despite the importance of explainable artificial intelligence (XAI) in medical domains. This paper seeks to address this limitation by introducing comprehensive annotations for WBC images. Through collaboration with pathologists, a thorough literature review, and manual inspection of microscopic images, we have identified 11 morphological attributes associated with the cell and its components (nucleus, cytoplasm, and granules). We then annotated ten thousand WBC images with these attributes. Moreover, we conduct experiments to predict these attributes from images, providing insights beyond basic WBC classification. As the first public dataset to offer such extensive annotations, we also illustrate specific applications that can benefit from our attribute annotations. Overall, our dataset paves the way for interpreting WBC recognition models, further advancing XAI in the fields of pathology and hematology.

Motivation & Objective

• Address the lack of explainable AI in white blood cell (WBC) image analysis by providing morphological annotations beyond basic cell type classification.
• Bridge the gap between clinical diagnostics and AI by incorporating pathologist-validated morphological attributes crucial for hematological diagnosis.
• Enable development of interpretable machine learning models in hematology by creating a large-scale, detailed, and publicly accessible dataset.
• Facilitate research in explainable AI (XAI) by providing ground-truth annotations for attributes such as vacuoles, chromatin density, and granule morphology.
• Support downstream applications such as disease diagnosis and model interpretability through a standardized, comprehensive dataset.

Proposed method

• Identified 11 morphological attributes through collaboration with pathologists, systematic literature review, and manual inspection of WBC images.
• Annotated 10,298 WBC images from the PBC dataset with all 11 attributes, resulting in 113,000 total labels (11 attributes × ~10.3k images).
• Organized attributes into four categories: overall cell, nucleus, cytoplasm, and granules, each grounded in clinical relevance and medical references.
• Trained and evaluated multiple deep learning models (VGG16, ResNet50, ViT-Base, ConvNeXt-Tiny) on attribute prediction using 5-fold cross-validation.
• Used standard image classification architectures fine-tuned for multi-class, multi-label attribute recognition with cross-entropy loss.
• Evaluated performance using precision, recall, and F1-score across all attributes, with results reported per attribute and model.

Experimental results

Research questions

• RQ1 Can deep learning models achieve high accuracy in predicting detailed morphological attributes of white blood cells from microscopic images?
• RQ2 How do different vision transformer and CNN architectures compare in predicting WBC morphological attributes such as granule color and chromatin density?
• RQ3 To what extent do model performance metrics vary across attributes with different visual complexity or clinical rarity?
• RQ4 Can the proposed dataset support the development of explainable AI systems in hematology by enabling model interpretation through attribute-level analysis?
• RQ5 What is the impact of data quality and annotation consistency on the reliability of attribute prediction in medical image analysis?

Key findings

• The WBCAtt dataset is the first publicly available resource to provide dense, multi-attribute annotations for white blood cells at scale, with 113,000 labels across 10,298 images.
• ViT-Base achieved the highest F1-score of 99.96% for the granule color attribute 'red', indicating near-perfect prediction capability for this well-defined feature.
• For the challenging attribute 'nucleus shape' with multiple subtypes (e.g., segmented-bilobed, unsegmented-indented), ViT-Base achieved an F1 of 78.78%, showing moderate performance on complex morphological distinctions.
• The CNXT-Tiny model achieved an F1 of 99.57% for 'granule color: nil', indicating strong performance on absence-based attributes.
• The attribute 'chromatin density: loosely' had the lowest F1 (74.06%) across models, suggesting that subtle chromatin patterns remain a challenge for current models.
• The dataset enables high-precision prediction for most attributes (e.g., 99.86% F1 for 'granule type: round'), demonstrating its utility for training robust, interpretable WBC recognition systems.