Plasma Cell Enumeration By Manual and Automated Methods to Establish a Standard Pictorial Reference

Background The diagnosis of plasma cell dyscrasias requires accurate, reliable enumeration of bone marrow plasma cell burden. This is typically assessed by manual visual enumeration - usually with CD138 immunohistochemistry, to identify plasma cells on bone marrow biopsies. However, this manual method is limited by its inherent subjectivity. Objectives This study sought to assess the performance of automated enumeration of plasma cell burden, using an open-access bioimage analysis software, QuPath (v.0.3.2), and to produce a pictorial representation of plasma cell burden to assist in routine bone marrow reporting. Methods We analysed CD-138 stained bone marrow biopsy samples collected from 59 patients at Princess Alexandra Hospital (PAH) between 2020 and 2022. Each case was independently assessed by two blinded operators (A.M. and J.T.) to produce a visual enumeration estimate. Additionally, from the lab information system (LIS), the reported plasma cell percentage for each case was also acquired. Each slide was then analysed using QuPath, with a cell-segmentation algorithm. Data analysis was performed using IBM SPSS (v.29.0.2.0). Results For the training phase, 12 slides were utilised - and four regions of interest (ROIs) were drawn within each slide, totalling 48 ROIs. Cells within these ROIs were manually counted (range: 117 - 572 cells) and plasma cell burden was calculated (range: 3.33% -100.00%). These ROIs were used to develop and refine the final algorithm for automated enumeration. Using intraclass correlation (ICC) analysis, the concordance between automated enumeration and the manually counted plasma cell burden (i.e., the gold standard) was 0.931 (95% CI: 0.874 - 0.961). For the testing phase, all 59 slides were used for manual and automated enumeration. Their reported plasma cell burden ranged from 2.5% to 100%. Using ICC analysis, the concordance between automated enumeration and the three reads of manual enumeration (by observers: A.M., J.T. and the reported plasma cell percentage) was 0.964 (95% CI: 0.930 - 0.980). The concordance between automated enumeration and the reported plasma cell percentage was moderate at 0.848 (95% CI: 0.578 - 0.930). Heterogeneity and weakness of CD138 staining significantly compromised automated enumeration, in 15 of the 59 cases. Excluding these cases (n=44), the ICC concordance between automated enumeration and the reported plasma cell percentage was calculated to be 0.963 (95% CI: 0.910 - 0.982). Our study also found that plasma cell burden from visual estimation was consistently higher than automated enumeration by ~13%. Consequently, in the whole testing cohort (n=59), 32% of patients who were formally reported to have myeloma did not meet the diagnostic threshold of ≥ 10% plasma cells on automated enumeration. This figure was lower, but still substantial at 15%, within the sub-group of cases with optimal CD138 staining (n=44). Using slides with good-quality CD138 staining, 12 representative images of plasma cell burden (<5%, 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%) were also compiled to form a standard pictorial reference. Conclusions We produced a pictorial representation of plasma cell burden to assist in routine bone marrow reporting - to the best of our knowledge, this is the first study that has harnessed the power of automated and manual enumeration to do so. We found that automated enumeration correlates well with manual counting of plasma cells, and the current accepted standard of gross visual enumeration. Heterogeneity and weakness of CD138 staining can significantly affect results from automated enumeration, and slides being subject to automated analysis need to be specifically optimised in this regard.