Development of an Artificial-Intelligence-Based Tool for Automated Assessment of Cellularity in Bone Marrow Biopsies in Ph-Negative Myeloproliferative Neoplasms.

Simple Summary: In recent years, artificial intelligence has been used increasingly in medical practice. Technological progress has made it possible to digitize the slides of histological preparations, allowing the use of image processing and AI technologies in surgical pathology. This potentially reduces variability and improves the uniformity of some histological evaluations, such as cellularity assessment in bone marrow biopsies, which is traditionally performed visually by expert human observers, resulting in inter-observer and intra-observer variability. In this work, we developed an accurate AI-based tool for the automated quantification of cellularity in BMB histology and compared its performances with cellularity estimates from five expert hematopathologists. Our results showed the robustness of our model across users and two different scanners for digitized image generation. The cellularity assessment in bone marrow biopsies (BMBs) for the diagnosis of Philadelphia chromosome (Ph)-negative myeloproliferative neoplasms (MPNs) is a key diagnostic feature and is usually performed by the human eyes through an optical microscope with consequent inter-observer and intra-observer variability. Thus, the use of an automated tool may reduce variability, improving the uniformity of the evaluation. The aim of this work is to develop an accurate AI-based tool for the automated quantification of cellularity in BMB histology. A total of 55 BMB histological slides, diagnosed as Ph- MPN between January 2018 and June 2023 from the archives of the Pathology Unit of University "Luigi Vanvitelli" in Naples (Italy), were scanned on Ventana DP200 or Epredia P1000 and exported as whole-slide images (WSIs). Fifteen BMBs were randomly selected to obtain a training set of AI-based tools. An expert pathologist and a trained resident performed annotations of hematopoietic tissue and adipose tissue, and annotations were exported as.tiff images and.png labels with two colors (black for hematopoietic tissue and yellow for adipose tissue). Subsequently, we developed a semantic segmentation model for hematopoietic tissue and adipose tissue. The remaining 40 BMBs were used for model verification. The performance of our model was compared with an evaluation of the cellularity of five expert hematopathologists and three trainees; we obtained an optimal concordance between our model and the expert pathologists' evaluation, with poorer concordance for trainees. There were no significant differences in cellularity assessments between two different scanners. [ABSTRACT FROM AUTHOR]