Leukemia is among the leading causes of cancer-related deaths worldwide, with an estimated 487,000 new cases and 305,000 deaths reported in 2022 (Bray et al., 2024). This disease disproportionately affects people under the age of 40, with children being particularly impacted (Creutzig et al., 2018). The diagnosis of leukemia requires an expert hematologist with extensive experience, capable of analyzing peripheral blood slides employing costly medical equipment. Furthermore, morphological examination remains a cornerstone of leukemia diagnosis, providing accessible, low-cost, and widely available insights into hematologic malignancies (Greer and Means, 2009). Given the prerequisites for the diagnosis and prognosis of leukemia, many remote regions in underdeveloped countries experienced higher mortality rates (Rugwizangoga et al., 2022). However, early detection of leukemia enables timely and accurate treatment, which can significantly improve survival rates. For leukemia diagnosis, hematologists examine the peripheral blood film (PBF) of patients, a process that is often tedious, error-prone, and requires significant expertise. The limited availability of experts and costly diagnostic resources hinders early, precise, and timely detection of leukemia, particularly in resource-constrained areas. Furthermore, leukemia treatment requires repetitive blood sample preparation and detailed analysis, requiring frequent access to experts. AI-driven solutions are vital to overcome the low accessibility of experts. Such a system should be able to work across multiple microscopes and digitization scenarios, and to augment hematologists, it should provide explainable information to hematologists. Unfortunately, existing leukemia datasets (Scotti, 2005, Labati et al., 2011, Rezatofighi and Soltanian-Zadeh, 2011, Matek et al., 2019, Kouzehkanan et al., 2022, Bodzas et al., 2023) suffer from many limitations, including the absence of localization annotations of white blood cells, the lack of associated morphological attributes, and a limited number of images (Gao et al., 2023). Furthermore, there is no established multi-task learning (MTL) (He et al., 2024, Xu et al., 2023, Koyuncu et al., 2020) approach that simultaneously detects white blood cells and predicts their morphological attributes.

In addition to these limitations, these supervised algorithms (Sun et al., 2021, Tian et al., 2019, Zhang et al., 2022, Ultralytics, 2021, Wollmann and Rohr, 2021) require every cell in each image to be annotated, adding a significant manual burden. In the practical scenario, a hematologist uses only a small portion of the acquired image to perform the required analysis and annotation. Thus majority of the cells remain unannotated. On the other hand, requiring annotation of complete blood films not only increases time and cost but also limits the number of blood films that could be annotated.1 Employing existing object detection strategies over sparsely annotated datasets results in subpar performance since these methods are unable to use the unlabeled parts of slides. Complete annotation of a blood film in the monolayer takes around 200 min almost eight times longer than the 25 min needed for sparse annotation.

To overcome dataset challenges, we present a large-scale WBC dataset named ‘large leukemia dataset’ (LLD), that consists of two subsets ‘LeukemiaAttri’,2 capturing cell localization, classification, and morphological attribute prediction tasks; and ‘Sparse-LeukemiaAttri’, for the object-detection and attribute prediction based on sparsely annotated dataset task. An important highlight of the LeukemiaAttri is that it has been collected through two microscopes from different price ranges, one high-cost (HCM) and one low-cost (LCM), at three magnifications (100x, 40x, 10x) using various sensors: a high-end camera for HCM, a mid-range camera for LCM, and a mobile phone camera for both microscopes. In this collection, using high-resolution HCM (100x), experienced hematologists annotated 10.3k WBC of 14 types, having 55k morphological labels from 2.4k images of several PBS leukemia patients. The data set collection procedure is illustrated in Fig. 2 and further details are in Section 3. In conjunction with these datasets, we propose two models named leukemia attribute detector (AttriDet) and sparse Leukemia attribute detector (SLA-Det). For the task of detection of white blood cells (WBCs) along with the prediction of morphological attributes, we integrated a lightweight head, named AttriHead, into the existing object detector. AttriHead employs multiscale features from the feature backbone used in object detection. This not only enhances computational efficiency but also enables the backbone to become more informative through MTL. For the sparsely-supervised object detection, we propose a novel algorithm called SLA-Det. Note that in our case, on average, only 20% of the image is annotated shown in Fig. 1. To effectively use the unannotated region information, SLA-Det incorporates region-guided masking and pseudo-label selection based on multiple criteria, ensuring robust model performance. We believe that these extensions align our dataset and framework more with clinical practices and enhance the reliability of automated leukemia detection. Our extensive experimental results and discussion validate the proposed ideas and framework. Our contributions are given below.

Large leukemia dataset (LLD) consists of LeukemiaAttri and Sparse-LeukemiaAttri. LeukemiaAttri: A large-scale, multi- domain dataset of 2.4K microscopic images captured using microscopes of varying costs, two cameras, and three resolutions (10x, 40x, 100x). Each image includes WBC localization and morphological attribute annotations for an explainable diagnosis. Sparse-LeukemiaAttri: A 40x FoV dataset with sparsely annotated training data (1546 images) and fully annotated test data (185 images).

AttriDet is a fully supervised detector designed to identify WBCs and their morphological attributes.

SLA-Det is a sparse WBC detector that enhances pseudo-label selection by incorporating problem-specific constraints, including size, objectness score, and class label entropy.

Key results: Using sparse annotations, SLA-Det can process slides at 40x resolutions while maintaining accuracy, unlike AttriDet, which only worked on 100x resolution; thus, it can process slides 4x faster than AttriDet. SLA-Det outperforms state-of-the-art sparse detection methods by 34.7 mAP@50, demonstrating superior efficiency and performance.

The work emphasizes domain diversity, diagnostic interpretability, and efficient sparse annotation strategies for real-world leukemia analysis.