Artificial Intelligence-Based Quantitative and Topological Analysis of Plasma Cell Burden: A New Tool for the Assessment of Myeloma

Background

Diagnosis, risk stratification and the indication for therapy in myeloma may depend on the assessment of the burden of plasma cells in bone marrow biopsies. However, the routine assessment of the burden of plasma cells is a subjective estimation with limited inter-operator reproducibility. The aim of the project was to develop a machine-learning approach for the identification, quantitation and topological analysis of plasma cells in clinical bone marrow trephine samples in patients with myeloma and its precursor conditions.

Methods

Bone marrow biopsies from 50 cases of newly diagnosed multiple myeloma were identified between April 2022 and July 2021 from the Department of Pathology, Oxford University Hospitals NHS Trust, United Kingdom. Routine histological preparation of the bone marrow biopsy specimens comprised serial sections stained with Hematoxylin and Eosin (H&E); immunostaining for MUM-1, a specific nuclear marker of plasma cells in the bone marrow, as well as hybridization probes for Kappa and Lambda light chains. Imaging of stained biopsies were acquired using routine clinical digital automation (Philips Digital Pathology). After de-identification of patient data, the images of whole sections of bone marrow were subjected to automated cell-segmentation analysis (Figure 1A low power, 1B high power). Cells identified by a green dot represent MUM-1-stained plasma cells and red dots represent other nucleated non-plasma cells identified by the hematoxylin counter-stain. The relative proportion of the bone marrow occupied by plasma cells was calculated through the identification of MUM-1 positive cells divided by the sum of MUM-1-positive cells plus the remainder cells identified by hematoxylin. Quality control was performed on the samples comparing automated cell detection with manual identification.

Results

Of the fifty cases examined, the mean age was 70 years with a 1.5:1 Male : Female sex ratio. The mean number of plasma cells identified per slide was 21,252 (95% CI - 15,327 - 27,176). The mean plasma cell percentage according to manual assessment was 37% (95% CI - 33-42%) whereas the automated assessment identified a mean plasma cell percentage of 23% (95% CI - 20-16%). Figure 1C displays the results from each of the 50 cases paired by manual and automated enumeration results. Thresholds for decision-making in the diagnosis and risk stratification of myeloma are shown in red.

Discussion

The identification and enumeration of the plasma cell burden in bone marrow biopsy samples is feasible in routinely collected clinical samples. These data demonstrate the ability of an automated analysis of existing clinical-grade histological images to generate accurate and reproducible assessments of plasma cell burden. Not only was the enumeration consistently exaggerated by manual estimation, but the variance of the data was reduced by automated assessment, compared to manual estimation (Figure 1C).

As the plasma cell percentage from bone marrow biopsies is used to aid more and more clinical decisions such as risk-stratification in smoldering myeloma or indeed to initiate therapy as part of the SLiM criteria (IMWG 2016) the need for rapid, reproducible data grows. For example all four cases where the plasma cell percentage was ≥60%, automated analysis reported a result well below this potential threshold to start treatment. Such an approach represents an opportunity to improve clinical workflows through automation and allow the use of more precise data to stratify risk and inform clinical decision-making. However, recalibration of results based on immunostains, as well as updated clinical significance of these data will be required prior to routine clinical application.

The manual enumeration of plasma cells appeared to be heavily influenced by the pattern of plasma cell distribution through the bone marrow. Beyond simple enumeration of plasma cell burden, the distribution of plasma cell infiltrates into the bone marrow has been implicated in the pathophysiology of myeloma. Across our cohort of 50 cases, we observe a wide variety of distribution of plasma cells that range from interstitial infiltrates, with only minor plasma cell clustering, to arrangements where aggregates of plasma cells dominate. These observations appear to be independent of plasma cell burden and the development of topological descriptors, as well as their relation of disease progression, continues

Royston:Perspectum Limited: Current Employment. Aberdeen:Ground Truth Labs Limited: Current Employment, Current equity holder in private company. Sirinukunwattana:Ground Truth Labs: Consultancy. Gooding:Bristol Meyers Squibb: Research Funding.