Acute myeloid leukemia (AML) is one of the most common and deadly hematological malignancies in the elderly, with a median age at diagnosis between 68 and 72 years. AML patients over 60 years old have worse survival outcomes than their younger counterparts. Therefore, early and accurate diagnosis and subtyping of AML are crucial for patients' treatment and prognosis. However, diagnosing and differentiating AML subtypes in bone marrow smear images is a complex and time-consuming task. Recent advancements in deep learning have opened up new possibilities for medical diagnostics, such as the evaluation of skin cancer and the assessment of tumors of unknown primary origin based on pathological images. These advancements not only improve the accuracy of clinical diagnosis but also reduce the influence of human factors, providing a more objective and consistent basis for auxiliary diagnosis.

On March 21, 2023, Prof. QIAN Pengxu and Prof. HUANG He from the First Affiliated Hospital, Zhejiang University School of Medicine co-published an article entitled AMLnet, A deep-learning pipeline for the differential diagnosis of acute myeloid leukemia from bone marrow smears in the journal Journal of Hematology & Oncology. This study proposed a deep learning algorithm called AMLnet for the differential diagnosis of AML subtypes from bone marrow smears. The algorithm demonstrates diagnostic capabilities comparable to senior experts and provides a visualization of significant regions in the bone marrow smears that contribute to the prediction results, offering valuable diagnostic assistance for clinicians.

In an effort to address this critical clinical issue, the research team conducted a retrospective study and constructed a large-scale, wide-ranging AML bone marrow smear database from 2010 to 2021 for deep learning model training and testing. AMLnet, a two-stage deep learning model, was developed to determine whether an image comes from a patient sample and predict different AML subtypes at the image level, providing the top-5 prediction probabilities. For patients with multiple bone marrow smears, the integrated model combines the prediction results from different smears to achieve a comprehensive diagnostic prediction at the patient level.

AMLnet achieved an area under the curve (AUC) of 0.885 at the image level and 0.921 at the patient level when distinguishing between nine AML subtypes in an independent test set. In comparison with clinicians, AMLnet demonstrated a performance level higher than junior experts and comparable to senior experts. This indicates the model's high potential in terms of diagnostic accuracy and robustness for AML. Furthermore, the model utilizes the Grad-CAM deep neural network interpretability method to generate significance heatmaps, helping pathologists better understand the model's decision-making process and ultimately improving the accuracy and credibility of the diagnosis.

The AMLnet has been trained to diagnose and differentiate AML subtypes, demonstrating its potential to assume a pivotal role in AML screening and early diagnosis, remarked Prof. QIAN Pengxu.  By utilizing AMLnet, not only can clinicians' workload be mitigated, but patients' quality of life can also experience improvement. This has significant clinical and social implications, especially in areas with uneven distribution of medical resources, as it furnishes valuable diagnostic assistance for medical practitioners.

More information: Ph.D. candidate YU Zebin and Dr. LI Jianhu are the co-first authors of this article. Prof. QIAN Pengxu and Prof. HUANG He are the co-corresponding authors of this article.

Source: The First Affiliated Hospital, Zhejiang University School of Medicine