Artificial intelligence and the blood film: Performance of the MC‐80 digital morphology analyzer in samples with neoplastic and reactive cell types

Implementing artificial intelligence‐based instruments in hematology laboratories requires evidence of efficiency in classifying pathological cells. In two‐Universities, we assessed the performance of the Mindray® MC‐80 for hematology patients with frequent leukemic and dysplastic cells.


| INTRODUCTION
The hematology laboratory has recently been enriched with tools based on artificial intelligence methods for the morphological analysis of digital images of peripheral blood cells.The possibility now exists of obtaining automated differential leukocyte counts, alternative and complementary to the manual microscopic count, not by analyzing physical-chemical cellular parameters, as occurs with cell counters based on flow cytometry, but by reproducing the optical and cerebral actions that allow cells to be recognized and classified based on morphological, structural and chromatic characteristics, starting from a thinly spread and stained drop of blood.This type of equipment, the first versions of which, developed in the early 1980s, had been set aside by the manufacturers because they were too slow and cumbersome for easy use, have experienced a real renaissance in recent years, linked to significant developments in the technology used and in artificial intelligence, and a considerable spread in many countries.
Most of the models operating in today's laboratories are based on an original method of digital technology initially developed in Sweden. 1 After the digital transformation of microscopic images, they provide pre-classification of cells located in different fields of the film based on measurable morphological properties, such as size, contour, shades of color, texture, and fine geometric properties such as elongation, eccentricity and compactness. 2The images are presented on a screen and separated into the observed cell populations.Expert operators are necessary at the end of the process to validate the results, classify any unclassified cells, and reclassify those identified incorrectly.These tools generally pre-classify normal leukocytes with excellent accuracy and, with more variable accuracy, pathological ones. 1 Among the main advantages they offer, in addition to the reduced visual fatigue of the operators, is their use for remote training and consultations, the possibility of archiving images effectively and, potentially, better standardization and harmonization intra-and inter-laboratory.
We had the opportunity to evaluate, in 2021-2022, in two Universities of Rome, an image analyzer dedicated to the differential leukocyte count of new and original conception, the MC-80 system (Mindray ® Biomedical Electronics Co. Ltd., Shenzhen, China).We report in this article the results of such evaluation on samples from patients with hematological disorders whose peripheral blood (PB) films had a high frequency of leukemic and dysplastic cells.

| Performance analysis
To assess the ability to recognize and count normal cells, we compared the performance of the MC-80 WBC differential count in terms of reproducibility and comparison with the microscope reference method.In particular, we investigated the correlation between postclassification and microscopic results using linear regression to calculate the correlation coefficient (r) and the determination coefficient (r 2 ) for the five standard classes of leukocytes and four categories of pathological cells present on the film.Furthermore, we evaluated the ability of the system to highlight and recognize the presence of pathological cells as regards both the ability to signal the presence of which are normally absent in the circulation (flagging, sensitivity, and predictive value of positivity) and to exclude their presence in the appropriate samples (specificity and predictive value of negatives).

| The MC-80 instrument
The MC-80 system is an automated digital cell morphology analyzer that automatically locates blood cells, takes photos, and pre-classifies the film's blood cells.The LabXpert software, as used in our laboratory, runs on computers running the Windows 10 operating system.
The microscopic part has quality objectives, magnification up to 100Â, automatic dripping of microscope oil, and a sensor image camera of advanced conception.Based on the information provided by the manufacturer, the system captures images in the film and then takes them to 20 different depths of the field.A multi-layer fusion technology proceeds with the reconstitution-fusion of the images obtained at different depths to faithfully reproduce the cellular details and the methods of visually focusing the cells and their structures under the microscope.
On this basis, the software pre-classifies the cells captured in the film, grouping the cells of the various classes based on the level of maturation and any atypia.This process occurs continuously, and the average speed equals a preliminary analysis of 60 slides every hour (with an inevitable variability depending on the leukocyte count and the possible presence of artifacts).The system can also proceed with the morphological characterization of erythrocyte abnormalities and platelet counts; however, our evaluation did not include these aspects.The storage capacity of the scanned images is vast (estimated to be approximately 80 000 samples).

| Sample collection and film preparation
We used 591 3-mL blood samples collected in K2-EDTA (BD Vacutainer™, Fisher Scientific, Segrate, Italy) at the Hematology Department of the Università Sapienza of Rome (USR), made available for the study after standard hematological routine testing.A vast proportion of samples from hematology patients with blood disorders and a few with viral infections were selected to include most types of cells that do not normally circulate in the peripheral blood (Table S1).
They were based on diagnostic (morphological, immunophenotypical, histopathological, genetic, and molecular) tests in agreement with the WHO Classification of Tumours of Haematopoietic and Lymphoid Tissues. 3 This study was carried out according to the Helsinki Declaration, with institutional approval by the USR and Università Cattolica del Sacro Cuore (UCS), Rome.Based on clinical information, samples were blindly retrieved from patients of the hematology clinics to obtain a high proportion of films with pathological cells and were anonymized.Clinical diagnoses were obtained retrospectively (after the end of the study) from anonymized hospital records.They included acute and chronic leukemias of different origins, myelodysplastic and myeloproliferative syndromes, severe non-Covid-19 viral infections, leukemic lymphoma and plasma cell neoplasm (Table S1).A graphical representation of the validation study design is shown in Figure S1.Blood cell counts were obtained using a Mindray BC-6800 Plus at the USR laboratory within 6 h of blood collection.We used the Mindray automated system SC-120 in the USR laboratory for slide preparation, fixation and staining within 8 h from blood collection of two films for each sample for analysis with the MC-80, labeled C and D. In addition, two additional blood films were manually spread for each sample, labeled A and B, and automatically fixed and stained using the standard technique of the USR lab (Aerospray ® Gram Stainer-Cytocentrifuge Model 7320) for the reference microscope method.One observers' group at the USR for microscope analysis
For digital analysis (films C and D), we used the SC-120 automated blood film maker and stainer with different reagents and protocols to assess and identify an optimal final film quality:  We performed a microscope reference differential on at least 200 WBC on each film (films A at the USR and films B at the UCS.In the case of discrepancies between the two manual differentials, a third evaluator from UCS was used as the arbitrator.

| Cell classification
For this study, besides all normal circulating cell classes (eosinophils, basophils, lymphocytes, monocytes) and reactive lymphocytes, we grouped mature segmented neutrophils and band cells in the category of "neutrophils" and promyelocytes, myelocytes, and metamyelocytes in the category of "immature granulocytes (IGs)."Furthermore, we included neoplastic (abnormal) lymphoid cells (circulating B lymphoma cells and myeloma plasma cells), abnormal promyelocytes, and all types of blasts within the category of "neoplastic cells."

| Statistics
We evaluated within-run imprecision using the ICSH method 4 with 10 times repeated analyses of two randomly-selected samples.Analysis of duplicates with linear regression on 441 samples was also used to identify samples and cells that the system could classify differently on replicate runs.Disagreements in the result pairs were identified and analyzed to identify the reasons for instrument inconsistencies on repeated analysis of the same film.
We assessed comparability and agreement with the CLSI H20-A2 visual reference method 5 (i.e., accuracy) of the post-classification MC-80 results using, respectively, correlation analysis with linear regression and Bland-Altman analysis of differences. 5,6The pass rate was calculated as the proportion of results inside the 95% limits (±1.96SD).Outliers were isolated and verified at the MC-80 screen and the microscope to identify the reason for differences in cell classification between reference and test methods.
The study of clinical sensitivity was carried out using matrix tables and classification of each slide without cells which are typically absent in circulating blood as true negative (TN) or false positive (FP), and those who had such cells as true positives (TPs) or false negative (FN).
Following the protocol H20-A2 of Clinical and Laboratory Standard Association (CLSI) H20-A2, 7 we thus calculated the sensitivity, specificity, efficiency and predictive value of positive and negative MC-80 results for each abnormal cell category.Efficiency, in particular, was calculated as the percentage of results which were true results, whether positive or negative.S2).The coefficients of variations were variable in the two samples and well in-line with state of the art for neutrophils, eosinophils, basophils, and monocytes described for cytometry-based analyzers, 8 althoughthere was an exceedingly high CV% of lymphocytes for sample C261, which had a very low lymphocyte absolute number.

| Duplicate analysis (two runs of each of 413 films)
To assess the linear relationship and the internal consistency of the new method on samples with very heterogeneous cellular composition, the pre-classification of individual cell classes on each film of the C series was also correlated with the corresponding pre-classification result on films of the D series.Twenty-seven samples from the original 438 films were not included for technical reasons.The total of 413 samples, according to the microscope reference method, comprised films with IG (283, or 68.5%, with a range from 0.5% to 52%); with reactive lymphocytes (353, or 85.5%, from 0.5% to 36.5%); with neoplastic cells (198, or 48%, from 0.5% to 97%); with nucleated red blood cells (NRBCs) (219, or 53%, from 0.5 to 446/100 WBCs).The results are expressed as the proportion of results comprised between the ±1.96SDs limits (i.e., pass rate, corresponding to the 95% central interval).Linear regression also provided us with coefficients of correlation (r), coefficients of determination (r 2 ), slope, and intercept (Table 1).The results confirm the excellent internal consistency of the MC-80 differential count, even in the presence of abnormal cells.
Importantly, a few outliers in the duplicate analysis of blood films were due to inconsistent classification of cells with highly abnormal nuclear or cytoplasmic features: These are described in detail below.

| Distributional inaccuracy
According to the CLSI document H20-A2, 7 the comparison of postclassification test method (MC-80) results with those of the reference method (optical microscope) using linear regression provides us with an estimation of comparability between instrument and microscope results, which in this case corresponds to a measure the accuracy of the test method.We have assessed both normal and abnormal cells.
Given the high reproducibility, we used the comparison slide C for the test method and the mean of 2 Â 200 counts (slides A and B) for the reference method (microscope).Three samples were excluded for technical reasons.
Table 2 displays the correlation coefficients (r) between MC-80 post-classification results and manual differential (reference) for the five types of normal leukocytes and the four primary populations of abnormal cells.We observed high r values for all types of cells except for reactive lymphocytes (0.437).
For completeness of information, we also compared the preclassification with the post-classification MC-80 results.Such results are reported in Supplemental Table 3. Differences between the major normal leukocyte classes were minimal, while the most relevant adjustments due to the operators' revision (post-classification) were observed for basophils and reactive lymphocytes.
The Bland-Altman scatter plot XY describes the agreement between two quantitative measurements. 5,6We calculated the Pass Rate according to the proportion of samples included in the interval between the limits of agreement, calculated as the interval ± 1.96 SD of the mean difference between the reference and the test method, corresponding to 95% of the data points.In the Bland-Altman plot, the X-axis corresponds to the mean of the two measurements, while the Y-axis corresponds to the difference between the two measurements (reference and test method) (Figure 1).Most discrepancies (i.e., points falling outside the 95% central interval) were due to slight variations.However, a minority of significant discrepancies existed for almost all cell types, which required searching for an explanation.

| Explanation of outliers (distributional disagreements)
Disagreements in a limited number of samples between the reference method and the MC-80 in neutrophils were due to the high frequency of dysplastic neutrophils in PB films obtained from patients of the Hematology Department of the ULS (Figure 2).Similarly, the high frequency of cells with abnormal chromatin condensation and cytoplasmic granulation was responsible for some disagreements in classifying IG.In the case of lymphocytes and blast cells, a significant disagreement was observed in nine samples, visible in the correlation graphs and Bland-Altman plots.In these nine samples, the pathological cells represented the dominant population on the films, obtained in almost all cases from patients with follicular lymphoma or splenic lymphoma with villous lymphocytes (splenic marginal zone lymphoma).Such lymphomatous cells were small lymphoid elements with thin cytoplasm and dense cleaved nuclei or villous cytoplasm, which were classified as lymphocytes by the MC-80 (even in post-classification) and lymphoma cells (included among neoplastic cells) by the observers at the microscope (Figure 2E).At the retrospective revision of these outliers, it was evident that the specific morphological details of these lymphoma cells were well preserved on the MC-80 screen.During the post-classification process, the reviewers confirmed the automated classification as lymphocytes (small dense nuclei) when reviewing slides at the MC-80 screen.In contrast, the reviewers had classified the cells as abnormal (neoplastic) at the microscope, likely thanks to those samples' lymphocytosis and cellular monomorphism.However, All nine samples were alerted by multiple flags from the blood cell counter and the digital analyzer so that a slide review by an expert morphologist (pathologist or hematologist) would have been granted in all cases.

| Clinical sensitivity: Identification of films with pathological cells
According to the H20-A2 standard, to evaluate the diagnostic performance of the MC-80 digital analyzer, we applied to the differential counts (digital machine vs. reference method) the Galen and Gambino matrix table method. 7,9The results obtained for the different classes of cells are shown in Table 3. Considering clinically relevant performance, we considered as positive for the respective cell type those samples with more than 1% IGs, more than 3% reactive lymphocytes, any percentage of neoplastic cells (blasts or circulating myeloma cells), and more than one NRBC per 100 leukocytes.
Regarding detecting IGs, only two samples out of 438 were FNs (with 1.5% and 2% IGs, respectively), so the sensitivity of the MC-80 results is extremely good for this type of cell.The predictive value of a negative MC-80 result for IG is 99.2%.On the other hand, specificity was less good, with 33 FP results and a predictive value of positive results equal to 83.5%.
The sensitivity of the MC-80 for neoplastic cells was 83.8%, with a predictive value of negative results of 88.4%.Among the 36 FN films, four showed 0.5% of neoplastic cells at the optical microscopy, three showed 1%, and one had 1.5% neoplastic cells at the microscope.Nine samples had a morphological blast percentage between 2% and 5% and three between 5% and 10%.In all the cases, neoplas-   system and as malignant, mature-looking lymphoma cells (manually included in the neoplastic cell count) by the microscope operator, and the reason for the misclassification was apparent (small size, dense chromatin, high nucleocytoplasmic ratio).The predictive value of a positive (>0%) MC-80 count for neoplastic cells was 80.4%, with 36 FP samples and a specificity of 86.1%: even in this case, the misclassifications were primarily due to the fact that several reactive lymphoid cells (at the microscope) were classified as neoplastic cells by the system and, rarely, vice versa.All these samples were flagged by the digital morphology analyzer and required an expert morphological review at the microscope in a clinical setting.
The MC-80 displays, on the other hand, a good efficiency (95.2%) and sensitivity (93.6%) for reactive lymphocytes, with a predictive value of negative results equal to 97.8%.As for NRBC, nine FN films out of 122 were positive for the reference method, in all cases with less than five NRBCs per 100/leukocytes.Sensitivity is 97.5%, with a predictive value of a negative result equal to 96.9%.

| DISCUSSION
Our study was specifically designed to subject the new digital morphological analyzer to extreme performative stress by double-blinded analysis of a large number of blood films from patients of one of the main hematology centers in Rome (USR).(Table S1).The instrument's location in a different, highly specialized university structure (UCS) guaranteed the total anonymity of the samples analyzed and the lack of clinical information.The reference leukocyte microscopic formula was also produced with the contribution of both laboratories, involving an additional element of safety and potential inhomogeneity of the collected data.These difficulties were successfully overcome thanks to careful planning and implementation of the experimental project.Under such conditions, the digital morphology analyzer we evaluated using recognized reference standards in a setting with a high incidence of films with pathological cells has shown adequate efficiency in identifying films with IGs, neoplastic cells, reactive lymphocytes, and NRBCs.We found an excellent correlation between the test and the reference method for normal and abnormal circulating cell types, except for reactive lymphocytes.Such a result was confirmed by the pass rate and the Bland-Altman analysis.
Given the setting with a high proportion of hematological abnormalities, the core of our study was the assessment of clinical performance in identifying films containing leukemic blasts, IGs or NRBCs.
Sensitivity, specificity, and predictive values for identifying films containing such cells which are normally absent in circulating blood, were good for IGs, NRBCs, neoplastic cells (blasts) and reactive lymphocytes.0][11][12][13] They ensure the excellent clinical performance of the MC-80 in terms of clinical diagnostic utility and as a completion of the WBC differential count provided by the flow cytometry-based blood cell counters.
We could identify a specific number of outliers due to differences in nomenclature, in particular for mature-looking lymphoma cells with dense chromatin and barely indented or cleaved nuclei, which were classified as abnormal lymphocytes (included in the blast count) at the microscope and as lymphocytes at the post-classification MC-80 differential count.In other samples, the bizarre nuclear shapes and abnormal cytoplasm granularity of neutrophils in films from patients with myelodysplastic syndromes or acute myeloid leukemia also caused their misclassification, more often as IGs or monocytes.
Inconsistencies in identifying lymphoid cells as normal or reactive lymphocytes were likely due to the presence, in several films, of elements with borderline morphology between the two cell classes and the heterogeneity of shapes and colors in the category of reactive lymphocytes.We a posteriori verified that the criteria for classifying lymphocytes as "reactive" were imperfectly harmonized between the different microscope observers and MC-80 post-classification operators, which explains the reduced quality of the results we obtained with this specific cell class.In future studies based on morphological cell identification, a better harmonization of criteria for classifying reactive lymphocytes will be necessary.Such issues were responsible for some inconsistencies in the classification of those last, both at the microscope and at the digital cell analysis. 14 The sensitivity for specimens containing blast cells was 83.8%, with a predictive value of negative results of 88.4% (Table 3).It must be underlined that our study aimed at identifying specific sensitivity for each abnormal cell class population, and, as already mentioned, a small number of malignant cells were included in non-blast categories, such as reactive lymphocytes and, rarely, IGs or monocytes.However, all samples containing neoplastic cells in any percentage at the microscope reference method were flagged by the automated blood cell counter or by the digital morphology analyzer, so a morphological revision in a clinical setting would have been necessary.
Among the limitations of our study, our test sample population did not include samples with pseudothrombocytopenia.We did not analyze a sufficient population of healthy subjects to obtain reference values (which were not among our primary targets).We did not evaluate the capacity of the system to identify and enumerate apoptotic cells and smudge cells, mainly owing to a problematic and potentially unreliable identification of such elements for the reference method at the optical microscope.In addition, state-of-the-art data for digital morphology analyzers are still unavailable in the literature.In particular, Vis and Huismans criterion is not suitable for digital image analyzers, which count only 200 cells.
Among the merits of the MC-80 system, we appreciated the excellent morphological reproducibility and similarity with microscope images of cells in normal and dysplastic samples (Figure S2).This quality is essential to strengthening the position of this new digital analyzer in all the areas that have classically been advocated as fields of application and exploitation of digital morphology, such as remote reviewing, training and education, harmonization and comparability of results, peer film reviewing and consultation. 1 kept the A set of films.The other three sets (B, C, and D) were transported weekly by car, within a 15-min drive, to the Transfusion Center of the Catholic University of Sacred Heart (UCS) of Rome.At UCS, the C and D sets of films were used to obtain duplicate 200-cell WBC differential counts with the digital analyzer Mindray MC-80 and the B set of films for second observers' optical microscopy.

2 . 2 . 5 |
Merck ® (Merck KGaA, Darmstadt, Germany) protocol (150 samples) Merck May-Grünwald eosin-methylene blue solution modified for microscopy: 0.5-min prefixation with methanol +3-min stain with May-Grünwald +8.5-min stain with Giemsa reagent.Blood film analysis C and D films were digitally processed in different batches at the TCL Laboratory on the MC-80.According to the UCS Laboratory standard, a hematologist or a clinical pathologist validated all pre-classification differential counts to obtain post-classification reportable results.

3 | RESULTS 3 . 1 |
Within-run precision (two samples, 10 runs) We measured the SD and CV of two abnormal samples analyzed 10 times each.The pre-classification differential count of the leukopenic sample C261 (WBC count 2.05 Â 10 9 /L) comprised 5.9% IGs, and the other sample C391 (WBC 7.34 Â 10 9 /l) comprised 0.3% of neoplastic cells and 3.5% of reactive lymphocytes.SDs of the replicate measurements were less than 2.5 (the instrument counts at least 200 cells per film) in all cases (Table tic cells were lymphoid or monocytic, and the differential count was flagged by the blood counter or the digital analyzer because of increased counts of reactive (atypical) lymphocytes or monocytosis.As mentioned before, nine additional films were obtained from patients with leukemic follicular lymphoma.The pathological cells, present in a high percentage (30%-98%) and flagged by the blood cell counter and by the MC-80, were classified as lymphocytes by the T A B L E 2 Comparison of the MC-80 post-classification results with the optical microscope reference method (Carlo Erba™ stain).
Examples of outliers in the duplicate analysis of films due to the inconsistent classification of highly abnormal cells.(A) In film C81 (red circle), some highly dysplastic hypogranular neutrophils were inconsistently classified as neutrophils or monocytes.(B) In film C21 (red circle), a case of acute promyelocytic leukemia under treatment, dysplastic granulocytes were inconsistently classified as neutrophils or abnormal promyelocytes (i.e., blasts).(C) In two films from patients with hyperleukocytosis (red circles) due to chronic lymphocytic leukemia, shrunk lymphocytes were inconsistently classified as lymphocytes or NRBCs.(D) In a case of acute promyelocytic leukemia (case 21, WBC 5.5 Â 10 9 /L), a dysplastic neutrophil with abnormal nuclear shape and chromatin is classified as an abnormal promyelocyte (i.e., in the blast category) (green circle).Note the correct automated classifications of the other hypergranular immature cells as abnormal promyelocytes.(E) In a case of leukemic splenic B-cell lymphoma (case 288, WBC 13.4 Â 10 9 /L), these mature-looking lymphoid cells with a round nucleus and frequently villous cytoplasm (green circles) were classified as lymphocytes by the MC-80 (pre-classification) and as lymphoma cells (i.e., in the blast category) by the observers at the microscope, due to their typical morphology.They also caused inconsistent lymphocyte counts in the duplicate imprecision assessment (31.9% in the first run vs. 67.5% in the second).In A-C, X-axis: first run: Y-axis: second run.Blasts in the Figure include neoplastic cells as lymphoblasts, myeloblasts, abnormal lymphocytes, abnormal promyelocytes and circulating myeloma plasma cells.Mon, monocytes; NRBCs, nucleated red blood cells.(Three samples not included for technical problems).
Similarly, mature B-cell lymphoma circulating cells with dense nuclear chromatin and scanty cytoplasm can be classified as "lymphocytes" by digital morphology devices: in general, such samples are flagged by the automated blood cell counter and the digital morphology analyzer.The pathologist or hematologist who reviews flagged slides with lymphocytosis should always pay special attention to morphological details that could represent morphological clues for a lymphoma diagnosis.As expected, we observed occasional inconsistencies of results and limited reproducibility in leukopenic films or cell categories present in low percentages (i.e., basophils and eosinophils, pathological cells), likely caused by their irregular distribution on the film.In rare cases, a few very small lymphocytes with extremely dark (pre-apoptotic) chromatin in hypercellular chronic lymphocytic leukemia were misclassified as NRBCs.Notably, most of the inconsistencies could be explained by specific factors, such as the cell's low number or very atypical shape (i.e., hyposegmented dysplastic neutrophils vs. IGs; lymphoma cells vs. blasts or normal to reactive lymphocytes).
Results of duplicate analysis of 413 films.
Note: Three samples not included for technical problems.Abbreviation: NRBCs, nucleated red blood cells.a Neutrophils include segmented and band neutrophils.b Immature granulocytes include metamyelocytes, myelocytes and promyelocytes.c Neoplastic cells include classified by the system as lymphoblasts, myeloblasts, abnormal (lymphomatous) lymphocytes, abnormal promyelocytes and circulating plasma cells in multiple myeloma patients.
Neoplastic cells include cells classified by the system as lymphoblasts, myeloblasts, abnormal (lymphomatous) lymphocytes, abnormal promyelocytes, and circulating plasma cells in multiple myeloma patients.
a Neutrophils include segmented and band neutrophils.bImmature granulocytes include metamyelocytes, myelocytes and promyelocytes.c Specific performance of the MC-80 in identifying films with specific types of cells that do not normally circulate in the blood of healthy subjects.
7offer state-of-the-art precision values for hematology analyzers that count thousands of cells, and maybe thisT A B L E 3 aNeoplastic cells include cells classified by the system as lymphoblasts, myeloblasts, abnormal (lymphomatous) lymphocytes, abnormal promyelocytes, and circulating plasma cells in multiple myeloma patients.Abbreviations: NRBCs, nucleated red blood cells; PV, predictive value.