An End-to-End Pipeline for Early Diagnosis of Acute Promyelocytic Leukemia Based on a Compact CNN Model

Abstract

Timely microscopy screening of peripheral blood smears is essential for the diagnosis of acute promyelocytic leukemia (APL) due to the occurrence of early death (ED) before or during the initial therapy. Screening manually is time-consuming and tedious, and may lead to missed diagnosis or misdiagnosis because of subjective bias. To address these problems, we develop a three-step pipeline to help in the early diagnosis of APL from peripheral blood smears. The entire pipeline consists of leukocytes focusing, cell classification and diagnostic opinions. As the key component of the pipeline, a compact classification model based on attention embedded convolutional neural network blocks is proposed to distinguish promyelocytes from normal leukocytes. The compact classification model is validated on both the combination of two public datasets, APL-Cytomorphology_LMU and APL-Cytomorphology_JHH, as well as the clinical dataset, to yield a precision of 96.53% and 99.20%, respectively. The results indicate that our model outperforms the other evaluated popular classification models owing to its better accuracy and smaller size. Furthermore, the entire pipeline is validated on realistic patient data. The proposed method promises to act as an assistant tool for APL diagnosis.

Digital assessment of peripheral blood and bone marrow aspirate smears

“…Rahman et al developed a CNN model trained on images of immature leukocytes capable of diagnosing acute myeloid leukemia cases with 96.5% accuracy 38 . By creating a three‐step pipeline consisting of leukocyte focusing, cell classification, and diagnosis modeling, Qiao et al developed an end‐to‐end pipeline for diagnosis of acute promyelocytic leukemia with 96%–99% precision 39 . Using CNN models, Ghaderzadeh et al were able to accurately differentiate hematogones and lymphoblasts representing early pre‐B ALL, pre‐B ALL, and pro‐B ALL 40 …”

Automatic generation of artificial images of leukocytes and leukemic cells using generative adversarial networks (syntheticcellgan)

“…The works in [5,6] are examples of CNN models proposed to classify the different normal leukocyte groups. Automatic recognition of acute leukemia has been mainly approached in two different ways: differentiating lymphoblasts and leukocytes [7,8] and classifying various lymphoblast lineages [9,10]. A CNN model was developed to distinguish neoplastic (leukemia) from non-neoplastic diseases (infections), as well as to recognize the lineage of leukemia [11].…”

A convolutional neural network‐based, quantitative complete blood count scattergram‐mapping framework promptly screens acute promyelocytic leukemia with high sensitivity

“…Recently, progress has been made in developing convolutional neural networks (CNNs) to recognize cell subtypes from microscopic images of peripheral blood or bone marrow smears. [12][13][14][15][16][17][18] In addition, there have been studies applying CNN to CBC scattergrams for other classification tasks. [19][20][21] Specifically, deep learning models have been applied to predict APL from cell images.…”

“…However, results from CBC tests do not have a one‐to‐one mapping relationship with any specific types of disease. Recently, progress has been made in developing convolutional neural networks (CNNs) to recognize cell subtypes from microscopic images of peripheral blood or bone marrow smears 12–18 . In addition, there have been studies applying CNN to CBC scattergrams for other classification tasks 19–21 .…”