Evaluation of a new technique using artificial intelligence for quantification of plasma cells on CD138 immunohistochemistry

Abstract

IntroductionThe diagnosis of plasma cell neoplasms depends on the accurate quantification of plasma cells, traditionally done by immunohistochemical CD138 staining of bone marrow biopsies. Currently, there is no fully satisfactory reference method for this quantification. In our previous study, we compared the commonly used overview estimation method (method A) with a novel method for counting plasma cells in three representative areas (method B). Results showed comparable concordance parameters between the two methods. In this follow‐up study, we compared the previously evaluated methods with a digital analysis method (method C) that uses artificial intelligence in open‐source software, QuPath.MethodsArchived CD138 immunohistochemically stained trephine sections of bone marrow samples used in our previous study were used (n = 33). Reviewers selected three representative areas on each sample by taking images with a light microscope and camera. Digital analysis was performed using the positive cell detection function in QuPath. The entire process was repeated by each reviewer to test intraobserver concordance (concordance correlation coefficient [CCC]) in addition to interobserver concordance (intraclass correlation coefficient [ICC]).ResultsIntraobserver concordance of method C showed strong agreement for all reviewers with the lowest CCC = 0.854. Interobserver concordance for method C using ICC was 0.909 and 0.949. This showed high interobserver agreement with significant differences between method C and previously assessed method A (ICC = 0.793 and 0.713) and method B (ICC = 0.657 and 0.658).ConclusionWe were able to successfully count CD138‐positive plasma cells in bone marrow biopsies using artificial intelligence. This method is superior to both manual counting and overview estimation, regardless of tumour load.