Automated myeloma cell selection using machine learning and artificial intelligence.

Abstract

Listen

Translate article

7565 Background: Precision of sampling is critical to achieve accurate results relevant to diagnostics and prognostics. Sampling errors may compromise the sensitivity and specificity of otherwise credible diagnostics technologies. Our 3- dimensional (3D) telomeres profiling methodology is conducted on individual target cells selected from the patient sample processed using 3D immuno-FISH. In our traditional workflow target cells are manually selected by highly trained personnel based on morphological characteristics and/ or immunophenotypic characteristics, namely cells positive for CD138 and CD56. However, the variability among trained operators required rounds of verification by a second and third operator and an approval by a Hematopathologist or the Lab Director to confirm the selected cells in some cases. Therefore, automated detection of myeloma cells is of paramount importance. Methods: In this study we employed machine learning and artificial intelligence (AI) tools to develop an automated algorithm capable of streamlining target cell (i.e. myeloma) selection and minimize or eliminate operator introduced variability. Training libraries were built using positive and negative cells. Over 5000 cells were used in the algorithm training process. We employed iterative rounds of validation and refinement to achieve a high precision automated tool. AI then identified myeloma cells in an automated fashion and the results of AI identification were confirmed by a Hematopathologist. Results: We report the results of a comparative analysis done on 20 myeloma patients at different stages of the disease including MGUS, smouldering myeloma and active myeloma. Also, the sample type included bone marrow smears, purified mono-nucleated plasma cells and bone marrow biopsies. 50 verified manually selected target cells and 50 target cells selected by the automated cell selection tool, approved by a skilled operator, from the same sample of each patient were analyzed using the TeloView platform. The TeloView platform quantifies 6 telomeric molecular and structural parameters. We conducted statistical analysis to calculate Coefficient of Variation (CV) between the results of each patient. We set the acceptable variability to < 15%. We achieved CVs of less than 10% among all patients included in the analysis. Conclusions: The automated cell selection tool streamlined this critical sampling process in our workflow minimizing operator-introduced variability, reducing the processing time to a fraction of the time required for manual cell selection and verification which enables high throughput for the 3D telomeres profiling while maintaining precision.