In Annals of clinical biochemistry ; h5-index 25.0
BACKGROUND : Mainstream haematology analysers (HAs) are reported to have low detection sensitivity for platelet clumps. In this study, a deep learning (DL) algorithm, convolutional neural network (CNN), was implemented to detect platelet clumps.
METHODS : Adenosine diphosphate (ADP) was used to induce platelet aggregation to mimic platelet clumps detected (PCD) samples. Six types of leukocyte scattergrams were collected from the Sysmex XN-10. Then, multiple CNNs were trained and validated by scattergrams in a fivefold cross-validation (CV) method. Finally, the CNN model with the best CV accuracy was tested with practical routine work samples.
RESULTS : A total of 386 samples (190 PCD and 196 negative samples) and 4,253 samples (150 PCD and 4103 negative samples) were eligible for CNN training and practical test, respectively. The CNN with the highest CV accuracy was trained by using scattergrams of side scatter (SSC) vs. forward scatter (FSC) from the white count and nucleated red blood cells (WNR) channel, whose mean area under the curve (AUC), accuracy, specificity, and sensitivity were 0.968, 0.940, 0.937, and 0.942, respectively, in the CV. In the practical test, the AUC, accuracy, specificity, and sensitivity of the CNN were 0.916, 0.961, 0.860, and 0.965, respectively. The dispersed spots presenting around the leucocytes in the WNR channel may be a sign of platelet clumping.
CONCLUSIONS : This study demonstrates that the CNN algorithms can identify platelet clumps based on optical information from dedicated leukocyte channels and has a higher ability to detect platelet clumps than the XN-10 device's internal algorithm under practical circumstances.
Fang Kui, Chen Xiling, Dong Zheqing, Xu Peng
2023-Jan-18
Laboratory methods, Quality assurance & control
