PCM-Leukemia: A Large Scale Dataset for Detection and Classification of Acute Leukemia Subtypes

Lin Tun Naing, Apichat Photi-A, Kaung Htet Cho, Matthew N. Dailey, Mongkol Ekpanyapong, Piya Rujkijyanont

Keywords: Acute Lymphoblastic Leukemia, Acute Myeloid Leukemia, Deep Learning, Object Detection, Classification, Domain Adaption

Abstract: Acute leukemia, consisting of acute lymphoblastic leukemia and acute myeloid leukemia, is a common hematologic malignancy. While the survival rate of patients with acute leukemia in high-income countries has significantly improved with contemporary chemotherapy regimens, the outcomes of those patients in low- and middle-income countries are still poor given delayed diagnosis and treatment. Although bone marrow examination is the gold standard for diagnosis of leukemia, detection of leukemia cells in peripheral blood can urge healthcare providers to start initial supportive care and refer patients to tertiary hospitals for definite diagnosis and proper treatment. Recently, Artificial Intelligence (AI) has shown promise in automating this process, yet the efficacy of deep learning models is often limited by the scarcity of large-scale, annotated datasets. To address this gap, we introduce a large-scale novel dataset, named PCM-leukemia, collected from Phramongkutklao Hospital and Phramongkutklao College of Medicine, comprising 19,191 images annotated by two hematology specialists and a trained biomedical researcher. The dataset includes bounding box and cell type annotations for eight distinct classes, including lymphoblasts and myeloblasts, yielding a comprehensive collection of 40,103 extracted single-cell crops. To validate the dataset’s utility for developing robust diagnostic tools, we established baselines using state-of-the-art object detection (YOLO11, DEIM) and classification pipelines. Specifically, we compared a standard CNN baseline (ResNet50) against a foundation model pretrained on histopathological images (DinoBloom), utilizing both linear probing and fine-tuning. Experimental results on our hold-out test set demonstrate the dataset’s high quality, supporting a strong mAP of 87.6% for WBC only detection and a classification accuracy of 92.59% with Macro F1-Score of 92.11% using the fine-tuned DinoBloom model. Furthermore, to assess the dataset’s capacity to facilitate generalization, models trained on our data were evaluated on external benchmarks for both ALL and AML subtypes. On the ALL-IDB1 dataset—re-annotated by our experts to include bounding boxes—the fine-tuned model demonstrated strong direct transferability, achieving an accuracy of 88.02% and a Macro F1-Score of 74.27% without training on the external set. Conversely, evaluation on the Munich AML Morphology Dataset (LMU) revealed a more challenging transfer scenario, yielding a baseline accuracy of 39.67% and Macro F1-Score of 40.63% in the direct transfer setting. To address this domain shift issue, we employed a low-resource supervised adaptation strategy; by incorporating just 10% of the target data into the training process alongside our proposed dataset, accuracy on the remaining 90% hold-out set increased significantly to 81.94%, and the Macro F1-Score increased to 67.94%. These results confirm that the proposed dataset captures representative features necessary for training generalizable and adaptable AI systems.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Integration of Imaging and Clinical Data

Registration Requirement: Yes

Reproducibility: codebase: https://github.com/ l-kuo/pcm-leukemia, dataset: https://qnap-2.aicenter.dynu.com/ share.cgi?ssid=bd169009b6d048c6bfa802043baa6601

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 43