TLDR: AttriGen is a novel AI framework that automates the detailed annotation of blood cell images, focusing on both cell type and 11 fine-grained morphological attributes. It uses a dual-model architecture, combining a CNN for cell type classification and a Vision Transformer for attribute recognition, achieving 94.62% accuracy. This system significantly reduces the time and cost associated with creating large, richly annotated medical datasets, accelerating research and diagnosis in hematology by providing comprehensive, clinically relevant characterizations of blood cells.
Diagnosing diseases like leukemia often relies on the meticulous microscopic review of blood smears, a process that is both time-consuming and labor-intensive, and further complicated by a shortage of skilled laboratory experts. While automated systems exist, most focus on basic cell type classification, frequently missing the subtle, fine-grained morphological details crucial for accurate clinical decisions.
A new framework called AttriGen aims to bridge this critical gap. Developed by Walid Houmaidi, Youssef Sabiri, Fatima Zahra Iguenfer, and Amine Abouaomar from Al Akhawayn University, AttriGen introduces an automated, fine-grained multi-attribute annotation system specifically designed for blood cell datasets. This innovative approach significantly enhances model interpretability and offers substantial time and cost efficiencies compared to traditional human annotation methods.
AttriGen employs a sophisticated dual-model architecture that mirrors the workflow of a human pathologist. It combines a Convolutional Neural Network (CNN) for classifying eight distinct blood cell types, trained on the Peripheral Blood Cell (PBC) dataset, with a Vision Transformer (ViT) for predicting eleven expert-defined morphological attributes. These attributes include cell size, cell shape, nucleus shape, nuclear-cytoplasmic ratio, chromatin density, cytoplasm texture, cytoplasm color, cytoplasm vacuole, granularity, granule type, and granule color. The ViT model is trained on the WBC Attribute Dataset (WBCAtt), which contains these detailed annotations.
The fused output from both models provides a comprehensive 12-attribute profile for each blood cell – the cell type plus eleven morphology traits – offering a richer, clinically aligned characterization. This system achieved a new benchmark of 94.62% accuracy in multi-attribute classification, demonstrating near human expert performance.
One of AttriGen’s most significant contributions is its automated annotation pipeline. This pipeline applies the dual model to unlabeled images, generating high-quality multi-attribute labels with minimal human intervention. This capability allows for rapid dataset expansion, accelerating the development of downstream models in medical imaging. For instance, annotating 6,784 unannotated cells in the PBC dataset, which would typically take weeks of extensive manual effort, can be completed in approximately 2.26 minutes using AttriGen.
The framework’s impact extends beyond just efficiency. It democratizes access to richly annotated datasets, enabling researchers without vast annotation resources to develop and validate attribute-based recognition systems. It also facilitates the exploration of complex relationships between morphological attributes and disease states that might otherwise remain unexamined due to data limitations. While initially demonstrated in hematological image analysis, the AttriGen approach is generalizable, establishing a paradigm for expanding attribute-based image classification datasets across diverse medical imaging domains, potentially accelerating the development of more interpretable and clinically relevant computer vision systems throughout healthcare.
The research paper, titled “AttriGen: Automated Multi-Attribute Annotation for Blood Cell Datasets,” can be found at arXiv:2509.26185.
Also Read:
- Enhancing Medical Image Analysis with Saliency-Guided AI for Longitudinal Studies
- Advancing Breast Cancer Diagnosis: A Transformer-Based Approach
Future research will explore extending AttriGen to pathological cells with abnormal morphologies, which are crucial for diagnosing conditions like leukemia. Additionally, integrating active learning techniques and enhanced explainability methods could further optimize the annotation process and highlight influential image regions for specific predictions.
