Evaluating a Machine Learning Approach for the Detection of AML

Researchers led by Jan-Niklas Eckardt, MD, of the Technical University Dresden in Germany, have developed a machine learning approach that allows for rapid and highly accurate identification of acute myeloid leukemia (AML) from bone marrow cytomorphology. According to Dr. Eckardt, this proof-of-concept model demonstrates the feasibility of image analysis, guided by so-called convolutional neural nets (CNNs), in the diagnosis of hematologic malignancies.

“Even for experts, diagnosis from image data alone is often challenging and diagnostic precision may suffer from significant inter-rater variability in edge cases,” Dr. Eckardt told ASH Clinical News. “This is especially true for the evaluation of bone marrow smears in the diagnostic process of hematologic malignancies.”

Dr. Eckardt noted that cytomorphology represents an essential component at the center of the diagnostic workflow. “Not only is cytomorphology ubiquitously available,” he said, “it’s cost-effective and can provide important information at a glance.” However, the process of acquiring a sample, staining it, counting cells under the microscope, and establishing a final diagnosis remains largely unchanged.

In the study, Dr. Eckardt and colleagues developed a CNN-based model capable of distinguishing between patients with AML and healthy individuals based on digital bone marrow images. The researchers trained the machine learning pipeline with samples from 1,251 patients with AML as well as a control group comprising 236 bone marrow samples from healthy donors. Samples were obtained from prior multicenter trials and the German Study Alliance Leukemia multicenter bioregistry.

The investigators trained a variety of CNN classifiers to detect individual cell types using supervised learning with a human-in-the-loop approach, according to Dr. Eckardt. For the CNN-based diagnosis of AML, the investigators achieved an area-under-the-receiver-operating-characteristic (AUROC) of 0.97 using only image data from bone marrow smears.

Additionally, the researchers used neural nets to predict the mutation status of NPM1, a common gene alteration in AML that also holds clinical implications for risk prediction. The CNNs accurately predicted NPM1 mutation status from bone marrow smear image data alone with an AUROC of 0.92. “With respect to explainable artificial intelligence, we used occlusion sensitivity maps which unveiled a so far unreported morphology of NPM1-mutated and -wildtype myeloblasts,” Dr. Eckardt added.

The machine learning pipeline is both scalable and transferable, “as cell detection and correct cell-type classification is essential in a wide variety of use-cases in bone marrow diagnostics,” Dr. Eckardt said. “While the standard diagnostic procedure may take hours or days until initial assessment, our algorithm merely requires an image from a bone marrow smear and provides a class prediction within seconds. Potentially, CNN-guided image analysis can aid in the diagnostic process by providing medical professionals with a fast and standardized pre-assessment for expert validation thereby increasing diagnostic precision and efficiency.”

He added that effective training of neural nets to improve accuracy in class predictions relies largely on cooperative efforts and shared datasets and source codes in order to build optimized models that can drive the application of artificial intelligence in medicine.

Additionally, large multicenter and multinational data sets may not only offer sufficient training data for classifiers, but also help to reduce bias.

“Computer vision holds the potential to change the diagnostic process in hematology providing accurate results in a shorter period of time,” Dr. Eckardt said. “Such algorithms, if trained and validated properly, may provide medical professionals with robust decision-making tools to guide treatment allocation. However, open access to data sets and source codes are essential for unbiased model building and prospective validation is needed to ensure patient safety.”

Study authors report no relevant conflicts of interest.

Reference

Eckardt JN, Middeke JM, Riechert S, et al. Deep learning detects acute myeloid leukemia andpredicts NPM1 mutation status from bone marrow smears [published online ahead of print,2021 Sep 8]. Leukemia. doi:10.1038/s41375-021-01408-w.