Automated diagnostic support system with deep learning algorithms for distinction of Philadelphia chromosome-negative myeloproliferative neoplasms using peripheral blood specimen

Abstract

Philadelphia chromosome-negative myeloproliferative neoplasms (Ph-negative MPNs) such as polycythemia vera (PV), essential thrombocythemia (ET), and primary myelofibrosis are characterized by abnormal proliferation of mature bone marrow cell lineages. Since various non-hematologic disorders can also cause leukocytosis, thrombocytosis and polycythemia, the detection of abnormal peripheral blood cells is essential for the diagnostic screening of Ph-negative MPNs. We sought to develop an automated diagnostic support system of Ph-negative MPNs. Our strategy was to combine the complete blood cell count and research parameters obtained by an automated hematology analyzer (Sysmex XN-9000) with morphological parameters that were extracted using a convolutional neural network deep learning system equipped with an Extreme Gradient Boosting (XGBoost)-based decision-making algorithm. The developed system showed promising performance in the differentiation of PV, ET, and MF with high accuracy when compared with those of the human diagnoses, namely: > 90% sensitivity and > 90% specificity. The calculated area under the curve of the ROC curves were 0.990, 0.967, and 0.974 for PV, ET, MF, respectively. This study is a step toward establishing a universal automated diagnostic system for all types of hematology disorders.

Applications of Artificial Intelligence in Philadelphia-Negative Myeloproliferative Neoplasms

“…Kimura et al [ 27 ] established a clinical decision support system for Ph-negative MPNs to minimize workload and inter- and intra-personal discrepancies ( Figure 2 ). The technique involved combining the complete blood counts (CBCs) and research data collected by an automated hematological analyzer (Sysmex XN-9000) with peripheral blood (PB) smear morphological features retrieved using a recently developed Convolutional Neural Network (CNN) coupled with an Extreme Gradient Boosting (XGBoost)-based decision-making algorithm [ 34 , 35 ].…”

Deep learning applications in visual data for benign and malignant hematologic conditions: a systematic review and visual glossary

“… 58 For certain disorders, the detection of aberrant WBC morphologies from peripheral blood is paramount. DL algorithms consistently detect dysplastic neutrophils pathognomonic for myelodysplastic syndrome (MDS), 59 as well as other white blood precursors to aid in the diagnosis of MPN, 60 acute promyelocytic leukemia (APL), 12 or acute lymphoblastic leukemia (ALL). 61 , 62 Many DL models for WBC detection have performed with high accuracy and AUROC upon internal validation strategies.…”

Is the interaction of technology useful in laboratory haematology diagnostics?

“…Very experienced operators are needed, because the morphology is complex and also because, otherwise, there is a high degree of variability between observers. Digital microscopy is increasingly used in diagnostic laboratories, even though the devices have a large footprint, are relatively expensive, and have limited slip parameters that must be adhered to generate accurate differentials [10] [11]. Flow cytometry gives the maximum guarantee of results by using specific markers in the recognition of abnormal cells or abnormal phenotypes.…”

“…Digital microscopy is increasingly used in diagnostic laboratories, even though the devices have a large footprint, are relatively expensive, and have limited slip parameters that must be adhered to generate accurate differentials [10] [11].…”