Related Experiment Video
Updated: Jul 11, 2026
09:31
Automated Quantification of Hematopoietic Cell – Stromal Cell Interactions in Histological Images of Undecalcified Bone
Published on: April 8, 2015
11.5K
Synthetic bone marrow images augment real samples in developing acute myeloid leukemia microscopy classification
Jan-Niklas Eckardt1,2, Ishan Srivastava3,4, Zizhe Wang5
1Department of Internal Medicine I, University Hospital Carl Gustav Carus, TUD Dresden University of Technology, Dresden, Germany. jan-niklas.eckardt@uniklinikum-dresden.de.
NPJ Digital Medicine
|March 22, 2025
Summary
Generative adversarial networks (GANs) create realistic synthetic bone marrow smear images. These synthetic images effectively train deep learning classifiers for diagnosing leukemia, even for rare conditions.
Area of Science:
- Medical imaging
- Artificial intelligence in hematology
Background:
- High-quality image data is crucial for training deep learning (DL) classifiers.
- Data sharing for medical images is often restricted due to privacy concerns.
Purpose of the Study:
- To evaluate the utility of generative adversarial networks (GANs) in synthesizing bone marrow smear (BMS) images for DL classifier training.
- To assess the performance of DL classifiers trained with synthetic BMS data.
Main Methods:
- Digitized bone marrow smear (BMS) images from patients with acute myeloid leukemia (AML), acute promyelocytic leukemia (APL), and stem cell donors.
- Generated synthetic BMS images using StyleGAN2-Ada.
- Conducted a blinded visual Turing test with hematologists to assess image quality.
- Trained and evaluated DL classifiers using real and synthetic BMS data.
Main Results:
- Hematologists achieved 63% accuracy in identifying synthetic images, indicating high realism.
- DL classifiers trained on real data achieved area under the receiver operating characteristic curves (AUROCs) of 0.99 for AML, APL, and donor classifications.
- Classifier performance remained above 0.95 even with incremental substitution of real data with synthetic samples.
- Synthetic data addition improved classifier performance for the rare disease APL.
Conclusions:
- Synthetic BMS data generated by GANs are suitable for training highly accurate image classifiers.
- GAN-synthesized data can help overcome data limitations in medical imaging AI development.
- This approach shows promise for improving diagnostic accuracy in hematology, especially for rare diseases.
