Abstract
Background
The UIMD PBIA (ANI CO., Suwon, Korea) is a newly developed automated digital image analyzer using innovative algorithms for the analysis of peripheral blood smears. We evaluated the accuracy and throughput of UIMD PBIA for the classification of white blood cells (WBCs).
Methods
A total of 29,605 cells in 242 clinical samples (192 samples with abnormal findings and 50 normal samples) were used to evaluate the classification accuracy and throughput of the UIMD PBIA. In addition, the total processing time for WBC classification by UIMD PBIA was measured to calculate the throughput.
Results
UIMD PBIA revealed outstanding performance for the identification of normal samples (99.0% accuracy) and five-part differentials (neutrophil, lymphocyte, monocyte, eosinophil, basophil, 99.2% accuracy). Misclassifications frequently occurred for immature granulocytes (83.6-93.9% accuracy), blasts (93.5% accuracy), and abnormal lymphocytes (81% accuracy). The pathogenic cells were likely to be misclassified into other classes of the same lineage. The average throughput was approximately 42 slides per hour. In cases with pancytopenia, the throughput was approximately 29 slides per hour.
Conclusions
The UIMD PBIA offers the most accurate results for WBC classification and the highest throughput, thereby reducing the technical workload, especially in cases with normal findings and pancytopenia. Accordingly, this study revealed the feasibility of using a digital switch in CBC analysis.
초록
배경
UIMD PBIA (ANI CO., 한국)는 새롭게 개발된 말초혈액 혈구 이미지의 자동화 분석장비이다. 본 연구에서는 UIMD PBIA의 백혈구 분류의 정확도와 처리속도를 평가하였다.
방법
이상소견이 있는 환자 검체 192건과 이상소견이 없는 정상인 검체 50건을 포함한 총 242건의 말초혈액 도말로부터 얻어진 29,605개의 백혈구 세포 이미지를 이용하여 장비의 정확도와 처리속도를 분석하였다.
Automated complete blood cell (CBC) analyzers have evolved to flag abnormal peripheral blood specimens. However, a manual slide review for the validation of the CBC test results is inevitable. The Clinical and Laboratory Standards Institute (CLSI) guideline requires a manual differential count of 200 cells performed by two experienced diagnostic hematologists [1]. This job is laborious and quality control is difficult to apply. Automated digital image analyzers have been developed to address these difficulties [2]. Since the development of the first image analysis system in 1966, remarkable progress has been made regarding its processing speed and convenience [2, 3]. The automated image analyzer was produced by Cellavision (CellaVision AB, Lund, Sweden) in the early 2000s. This system is integrated into the DI-60 platform (Sysmex, Kobe, Japan), which is equipped with a hematology analyzer and slide-making devices. The performance of the Cellavision and DI-60 has been thoroughly evaluated in recent studies [3-10]. Prior studies revealed suboptimal accuracy in the identification of abnormal cells, especially blasts and immature granulocytes. According to a recent paper by the International Council for Standardization in Haematology (ICSH), skilled morphologists must validate the automated results of classification using these devices. The paper also suggested that new instruments should possess improved accuracy for the classification of pathological cell types [9]. The UIMD PBIA (ANI CO., Suwon, Korea) is a novel automated digital image analyzer that uses innovative algorithms to review peripheral blood smears. This machine provides differential counts of white blood cells (WBCs) and morphological grading scores of red blood cells (RBCs) and platelets. Up to 12 stained slide glasses can be loaded into the input cassette, which is followed by continuous scanning of their barcodes and cells. This instrument provides high quality images and accurate differentials in a short time. In this study, we focused on the accuracy of the UIMD PBIA in the classification of WBC using heterogeneous clinical samples. In addition, we determined the throughput of the UIMD PBIA to establish its feasibility for clinical applications.
The automated digital image analyzer, UIMD PBIA, is equipped with a slide cassette, a slide barcode scanner, a microscope with two objectives (OPYMPUS UPLXAPO 10x and PLXN 100x), a camera (SONY XCL-SG510C, 2,464×2,056 pixel), and a computer system for acquisition and classification software (UPBA-12A, version1.0). The slide cassette can load up to 12 slides at a time. The graphical user interface is equipped with an operating system setting, real-time analysis monitoring, data storage, classification results, and reports. The captured WBC images are pre-classified into 13 types of cells: neutrophils, metamyelocytes, myelocytes, promyelocytes, lymphocytes, abnormal lymphocytes, monocytes, eosinophils, basophils, blasts, plasma cells, nucleated RBCs, and reactive lymphocytes. A manual reclassification is available using the “drag and drop” function on the screen. Both absolute count and percentages are presented in the final report. The nucleated RBC and artifacts (smudge cells) are counted separately (Supplementary Fig. 1).
A total of 242 clinical samples submitted for routine CBC analysis at the Seoul St Mary’s hospital were selected and sorted based on the result of conventional differential count. The classification accuracy was determined by analyzing 192 samples of 11 sample types, including nucleated erythrocytes (N=60), leukopenia (N=20), neoplastic dysgranulopoiesis (N=10), non-neoplastic dysgranulopoiesis (N=10), neutrophil precursors ≥10% (N=10), neutrophil precursors ≥10% (N=10), and abnormal cells such as myeloblasts and monoblasts (N=20) lymphoblasts (N=15), mature B cell neoplasm (N=15), mature T cell neoplasm (N=7), and plasma cell and/or reactive lymphocytes (N=15). Fifty samples without abnormal findings were also analyzed to evaluate the negative predictive power. The blood samples collected in anticoagulant tubes (BD vacutainer spray-coated K2EDTA tubes, Becton Dickinson, USA) were processed according to a standardized protocol. Blood films were prepared using slide maker and stainer (SP-10, Sysmex, Kobe, Japan) in an automated hematology slide preparation unit.
This study was carried out in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Seoul St. Mary’s Hospital (KC16DISI0316). The IRB waived the requirement of informed consent from patients as de-identified PB slides were used for the routine analysis.
The clinical evaluation was performed to assess the performance and reliability of the UIMD PBIA in normal and abnormal cases. Accordingly, at least 100 WBCs in each slide, captured and pre-classified by UIMD PBIA, were subsequently reviewed and corrected independently by two morphologists with over 30 years of experience in blood morphology examination. Discordant classifications between the two morphologists were discussed to obtain the final results. Manual curation was achieved with reference to the results of manual microscopic observation. Manual differential counts (200 cells) using light microscopy based on CLSI H20-A2 [1] were independently obtained by the two morphologists. To determine the accuracy of WBC classification, the average values of the two post-classifications by the two morphologists were compared with the pre-classification results. The accuracy of the classification was determined using 12 sample types (normal and abnormal samples) and 13 cell types. The throughput was determined by measuring the total processing time for WBC classification by UIMD PBIA.
The WBC pre-classification using the UIMD PBIA was compared with the post-classification levels to determine the classification accuracy. Overall, 97.1% of the pre-classification by UIMD PBIA for 29,065 cells was consistent with the reclassification. The UIMD PBIA had the highest accuracy for normal samples (99.0%), followed by samples with pancytopenia (97.7%), lymphoblasts (97.6%), mature T cell neoplasm (97.4%), neutrophil precursor more than 10% (97.3%), mature B cell neoplasm (97.3%), neoplastic dysgranulopoiesis (97.2%), neutrophil precursor less than 10% (97.0%), nucleated RBC (96.3%), non-neoplastic dysgranulopoiesis (96.3%), myeloblasts and/or monoblasts (95.9%), and plasma cells and/or reactive lymphocytes (95.6%) (Table 1).
The UIMD PBIA displayed the maximum accuracy for the pre-classification of neutrophils (99.9%), followed by eosinophils (99.8%), lymphocytes (99.1%), nucleated RBCs (97.0%), monocytes (96.6%), plasma cells (96.3%), myelocytes (93.9%), promyelocytes (93.9%), reactive lymphocytes (93.6%), blasts (93.5%), basophils (91.6%), metamyelocytes (83.6%), and abnormal lymphocytes (81%) (Table 2).
Table 3 presents details of the pre-classification results and corrected results by experts in the other categories. The most frequent misclassifications occurred for abnormal lymphocytes and metamyelocytes. Approximately 13.9% (11/79) of abnormal lymphocytes were reclassified into blasts and approximately 13.4% (112/836) of metamyelocytes were reclassified into neutrophils. Conversely, approximately 13.6% (12/88) of abnormal lymphocytes were misclassified into blasts, and lymphoblasts were frequently misclassified into lymphocytes (0.7%, 17/2,381), reactive lymphocytes (0.7%, 17/2,381), and abnormal lymphocytes (0.5%, 11/2,381) by UIMD PBIA. The median WBC counts per slide captured by UIMD PBIA was 121 (95% CI: 119–123). The classification accuracies were not associated with the cell counts (correlation coefficient r=-0.019, P=0.768).
The throughput of UIMD PBIA were measured for samples with normal findings (N=50), nucleated RBCs (N=60), and other findings (N=192). The total processing time was measured and divided by the number of slides and cells. Approximately 5 hours and 43 minutes were required to process 242 slides, suggesting a run time of 85 seconds per slide. The average throughput was approximately 42 slides per hour. Based on sample type, samples with normal findings had the best throughput (44.72 slides/hour), followed by other findings (41.67 slides/hour), nucleated RBCs (39.55 slides/hour), and pancytopenia (29.02 slides/hour) (Table 4).
Remarkably, the UIMD PBIA classified WBCs with an accuracy of 97%. Of note, the overall accuracy of the market leaders, Sysmex DI-60 and CellaVision DM96, ranges from 82% to 89% [4-8]. The UIMD PBIA displayed outstanding performance in the analysis of normal samples (99.0% accuracy) and five-part differentials (99.2%). According to the results, UIMD PBIA can replace the manual slide review process in the clinical laboratory with a low incidence of abnormal samples (Table 5).
Misclassifications were relatively high for immature granulocytes (83.6–93.9%), blasts (93.5%), and abnormal lymphocytes (81%). Nevertheless, UIMD PBIA had the lowest frequency of misclassification compared with other instruments (Table 5). UIMD PBIA has improved classification accuracy as it utilizes self-learning artificial intelligence. More than 300,000 PB slides from the real world were used to develop the classification algorithm. However, pathogenic cells were still likely to be misclassified into another class in the same lineage. Notably, this is a common phenomenon of image analyzers; thus, a careful review of blasts and abnormal lymphocytes is warranted [7, 8].
The UIMD PBIA distinguishes abnormal lymphocytes from atypical lymphocytes, consistent with reactive lymphocytes as suggested by the ICSH guideline [11]. To the best of our knowledge, this study included the largest number of samples with abnormal lymphoid cells (mature B cell neoplasm, N=15, mature T cell neoplasm, N=7) and plasma cells (N=15), and is the first report of classification accuracies involving abnormal lymphoid cells and plasma cells using a digital image analyzer without additional image processing [12-14]. The classification accuracy of plasma cells was acceptable (96.3%); however, as described previously, the abnormal lymphocytes were still insufficient (81%).
The UIMD PBIA yielded a significantly higher throughput than its competition and processed 42 slides per hour [7, 9, 15]. In the case of pancytopenia, UIMD PBIA processed 29 slides per hour. The cell tracking system minimizes the time required to search cells on the slide using an efficient path-finding algorithm.
Despite the commercial availability of high-performing automated digital image analyzers, these devices are supportive rather than an alternative to the microscopic slide review system and are associated with limited accuracy and throughput.
This study comprehensively investigated the newly developed digital image analyzer, UIMD PBIA, which offers the most accurate WBC classification results with the highest throughput. This improvement reduces the workload of the morphologist, especially in cases involving normal findings and pancytopenia [16]. Overall, this study demonstrates the feasibility of using the digital image analyzer, UIMD PBIA, as a digital switch for CBC testing.
REFERENCES
1. Clinical, Laboratory Standards Institute. 2007. Reference leukocytes (WBC) differential count (proportional) and evaluation of instrumental methods; Approved standard - Second edition. CLSI Document H20-A2. Wayne, PA: Clinical and Laboratory Standards Institute.
2. Da Costa L. 2015; Digital image analysis of blood cells. Clin Lab Med. 35:105–22. DOI: 10.1016/j.cll.2014.10.005. PMID: 25676375.
3. Tatsumi N, Pierre RV. 2002; Automated image processing: past, present, and future of blood cell morphology identification. Clin Lab Med. 22:299–315. DOI: 10.1016/S0272-2712(03)00076-3. PMID: 11933580.
4. Kratz A, Bengtsson HI, Casey JE, Keefe JM, Beatrice GH, Grzybek DY, et al. 2005; Performance evaluation of the CellaVision DM96 system: WBC differentials by automated digital image analysis supported by an artificial neural network. Am J Clin Pathol. 124:770–81. DOI: 10.1309/XMB9K0J41LHLATAY. PMID: 16203273.
5. Ceelie H, Dinkelaar RB, van Gelder W. 2007; Examination of peripheral blood films using automated microscopy; evaluation of Diffmaster Octavia and Cellavision DM96. J Clin Pathol. 60:72–9. DOI: 10.1136/jcp.2005.035402. PMID: 16698955. PMCID: PMC1860603.
6. Briggs C, Longair I, Slavik M, Thwaite K, Mills R, Thavaraja V, et al. 2009; Can automated blood film analysis replace the manual differential? An evaluation of the CellaVision DM96 automated image analysis system. Int J Lab Hematol. 31:48–60. DOI: 10.1111/j.1751-553X.2007.01002.x. PMID: 18177438.
7. Tabe Y, Yamamoto T, Maenou I, Nakai R, Idei M, Horii T, et al. 2015; Performance evaluation of the digital cell imaging analyzer DI-60 integrated into the fully automated Sysmex XN hematology analyzer system. Clin Chem Lab Med. 53:281–9. DOI: 10.1515/cclm-2014-0445. PMID: 25153399.
8. Kim HN, Hur M, Kim H, Kim SW, Moon HW, Yun YM. 2017; Performance of automated digital cell imaging analyzer Sysmex DI-60. Clin Chem Lab Med. 56:94–102. DOI: 10.1515/cclm-2017-0132. PMID: 28672770.
9. Kratz A, Lee SH, Zini G, Riedl JA, Hur M, Machin S, et al. 2019; Digital morphology analyzers in hematology: ICSH review and recommendations. Int J Lab Hematol. 41:437–47. DOI: 10.1111/ijlh.13042. PMID: 31046197.
10. Fu X, Fu M, Li Q, Peng X, Lu J, Fang F, et al. 2020; Morphogo: an automatic bone marrow cell classification system on digital images analyzed by artificial intelligence. Acta Cytol. 64:588–96. DOI: 10.1159/000509524. PMID: 32721953.
11. Palmer L, Briggs C, McFadden S, Zini G, Burthem J, Rozenberg G, et al. 2015; ICSH recommendations for the standardization of nomenclature and grading of peripheral blood cell morphological features. Int J Lab Hematol. 37:287–303. DOI: 10.1111/ijlh.12327. PMID: 25728865.
12. Alférez E, Merino A, Mujica L, Rodellar J. 2011; Morphological features using digital image processing in lymphoid neoplasias. Int J Lab Hematol. 33(suppl 1):53.
13. Alférez S, Merino A, Bigorra L, Mujica L, Ruiz M, Rodellar J. 2015; Automatic recognition of atypical lymphoid cells from peripheral blood by digital image analysis. Am J Clin Pathol. 143:168–76. DOI: 10.1309/AJCP78IFSTOGZZJN. PMID: 25596242.
14. Alférez S, Merino A, Mujica LE, Ruiz M, Bigorra L, Rodellar J. 2014; Automatic classification of atypical lymphoid B cells using digital blood image processing. Int J Lab Hematol. 36:472–80. DOI: 10.1111/ijlh.12175. PMID: 24325784.
15. Bengtsson HI. 2020; Digital morphology analyzers in hematology: Comments on the ICSH review and recommendations. Int J Lab Hematol. 42:e213–5. DOI: 10.1111/ijlh.13181. PMID: 32150335. PMCID: PMC7586827.
16. Kim AH, Lee W, Kim M, Kim Y, Han K. 2014; White blood cell differential counts in severely leukopenic samples: a comparative analysis of different solutions available in modern laboratory hematology. Blood Res. 49:120–6. DOI: 10.5045/br.2014.49.2.120. PMID: 25025014. PMCID: PMC4090333.
Table 1
WBC pre-classification accuracy of the UIMD PBIA based on sample type
Sample type | Sample characteristics | Number of samples | Cell counts | Accuracy | Appendix* |
---|---|---|---|---|---|
Total samples | - | 242 | 29,605 | 97.1% | |
Samples with NRBC | Nucleated RBC ≥ 5% | 60 | 7,568 | 96.3% | 1–1 |
Samples with pancytopenia | WBC < 1×109/L | 20 | 2,481 | 97.7% | 1–2 |
Samples with dysgranulopoiesis | Neoplastic dysgranulopoiesis (MDS, MPN/MDS) | 10 | 1,276 | 97.2% | 1–3 |
Non-neoplastic dysgranulopoiesis (post-chemotherapy) | 10 | 1,230 | 96.3% | 1–4 | |
Samples containing neutrophil precursors | Neutrophil precursors < 10% | 10 | 1,187 | 97.0% | 1–5 |
Neutrophil precursors ≥ 10% | 10 | 1,148 | 97.3% | 1–6 | |
Samples containing abnormal cells | Myeloblasts and/or monoblasts | 20 | 2,552 | 95.9% | 1–7 |
Lymphoblasts | 15 | 1,871 | 97.6% | 1–8 | |
Mature B cell neoplasm | 15 | 1,764 | 97.3% | 1–9 | |
Mature T cell neoplasm | 7 | 1,109 | 97.4% | 1–10 | |
Plasma cell and/or reactive lymphocytes | 15 | 1,783 | 95.6% | 1–11 | |
Normal samples | - | 50 | 5,532 | 99.0% | 1–12 |
*The appendix with details of the results is presented in the supplementary Table 1.
Table 2
WBC pre-classification accuracy of the UIMD PBIA based on cell type
Table 3
Details of the pre- and post-classification results
Table 4
Throughput of the UIMD PBIA
Sample types | Total Time* | Number of slides | Second/Slide | Slides/Hour | Number of cells | Second/Cell |
---|---|---|---|---|---|---|
Total | 05:43:30 | 242 | 85.16 | 42.27 | 29,605 | 0.69 |
Normal samples | 01:07:05 | 50 | 80.50 | 44.72 | 5,532 | 0.72 |
Abnormal samples | 04:36:25 | 192 | 86.38 | 41.67 | 24,073 | 0.68 |
Samples with nRBC | 01:31:02 | 60 | 91.03 | 39.55 | 7,568 | 0.72 |
Pancytopenia samples | 00:41:20 | 20 | 124.04 | 29.02 | 2,755 | 0.90 |
Table 5
Comparative accuracy of digital image analyzers
Sample types | Octavia [5] | DM96 [5] | DM96 [6] | DM96 [4] | DI-60 [7] | DI-60 [8] | UIMD PBIA |
---|---|---|---|---|---|---|---|
Segmented neutrophils | 94.4% | 98.6% | 99.5% | 92.5% | 98.9% | 96.6% | 99.9% |
Band neutrophils | 10.5% | 22.9% | 57.1% | 14.1% | 63.0% | ||
Eosinophils | 95.4% | 93.5% | 79.9% | 63.2% | 67.8% | 39.6% | 99.8% |
Basophils | 58.4% | 84.7% | 54.1% | 80.0% | 41.1% | 87.0% | 91.6% |
Lymphocytes | 94.3% | 95.2% | 94.9% | 96.4% | 86.6% | 88.9% | 99.1% |
Atypical lymphocytes | 70.7% | 93.6% | |||||
Abnormal lymphocytes | 81.0% | ||||||
Plasma cells | 96.3% | ||||||
Monocytes | 65.0% | 94.0% | 87.6% | 81.4% | 88.7% | 66.3% | 96.6% |
Blast cells | 84.4% | 78.5% | 76.6% | 65.1% | 93.1% | 56.0% | 93.5% |
Metamyelocytes | 32.6% | 53.2% | 23.6% | 33.0% | 83.6% | ||
Myelocytes | 37.7% | 30.6% | 75.0% | 93.9% | |||
Promyelocytes | 77.6% | 30.6% | 88.5% | 93.9% | |||
Nucleated RBCs | 89.6% | 86.7% | 30.6% | 94.4% | 97.0% | ||
Five differentials | 87.2% | 87.9% | 88.4% | 87.9% | 99.2% | ||
All | 87.0% | 92.0% | 89.2% | 82.0% | 87.6% | 86.0% | 97.1% |