Abstract
Leukemia diagnosis relies on the labor-intensive morphological analysis of white blood cells (WBCs). This paper introduces the Large Leukemia Dataset (LLD), a multi-domain benchmark, and SLA-Det, a novel detection framework designed to operate under sparse supervision. By leveraging pseudo-labeling and feature similarity, SLA-Det achieves high diagnostic accuracy with only 20% annotation coverage, while providing human-interpretable morphological descriptions.
Introduction
Leukemia remains a leading cause of cancer-related mortality, requiring hematologists to meticulously examine peripheral blood films (PBF) to identify and assess white blood cells. This process is tedious and requires specialized expertise. While deep learning has shown promise, it is often hindered by data scarcity, the "domain shift" between different hardware setups, and the immense time required for full image annotation.
The Large Leukemia Dataset (LLD)
To overcome these hurdles, the authors present the Large Leukemia Dataset (LLD), a robust benchmark that captures the diversity of clinical environments.
The LLD includes:
- LeukemiaAttri: A collection of 28.9K images from 48 patients, featuring 88,294 WBCs across 14 classes. It incorporates data from high-cost and low-cost microscopes, as well as mobile cameras, to ensure hardware agnosticism.
- Sparse-LeukemiaAttri: A subset where only 20% of each image is annotated, simulating the realistic scenario where clinicians only have time to label representative patches.
Methodology
The proposed framework consists of two primary innovations: AttriDet and SLA-Det.
AttriDet: Attribute-Aware Detection
AttriDet extends the YOLOv5 architecture with an "AttriHead" that predicts seven morphological attributes—such as cell size, nuclear shape, and cytoplasmic basophilia—alongside the cell class. This multi-task learning approach provides a regularization effect, improving the backbone's ability to recognize complex cell structures.
SLA-Det: Learning from Sparse Annotations
The Sparse Leukemia Attribute Detector (SLA-Det) is designed to utilize the unannotated 80% of images that standard models typically ignore.
SLA-Det employs two main strategies:
- Pseudo-Label Filtering: It identifies high-confidence predictions in unannotated regions to serve as additional training signals.
- Similarity-based Triplet Loss: It ensures that cells of the same type have similar feature signatures, whether they were officially labeled or identified via pseudo-labels.
Key Results and Discussion
SLA-Det significantly outperforms existing sparse detectors, achieving 45.1 mAP@50 on the sparse dataset—a 34.7-point lead over previous state-of-the-art methods. Furthermore, the model provides explainability by generating text-based descriptions that match hematological vocabulary, such as "blast cells with scanty cytoplasm and a high N:C ratio."
The system is also highly efficient, processing images in approximately 0.29 seconds, making it viable for real-time clinical assistance. While challenges remain in detecting rare features like cytoplasmic vacuoles, the framework represents a major step toward practical, frugal AI diagnostics.
Conclusion
By integrating a massive, diverse dataset with a clever sparse-learning strategy, this work demonstrates that high-performance leukemia diagnosis is possible without exhaustive manual labeling. The combination of hardware agnosticism and explainable outputs makes this approach particularly suitable for deployment in resource-constrained clinical settings.
Summary updated