Developing a Pathomics Imaging Technique to Assess Bone Marrow in MPNs

Among patients with BRC-ABL-negative myeloproliferative neoplasms, such as polycythemia vera (PV) and myelofibrosis, bone marrow (BM) evaluation is necessary for differentiating and classifying malignancies. A pilot study, presented at the 15^th International Congress on Myeloproliferative Neoplasms in Brooklyn, New York, sought to develop a deep learning (DL) technique that could estimate BM cellularity, characterize segmented nuclei, and identify variations in the BM microenvironment without the need for biopsies.

The study’s contributors, led by Spencer Krichevsky, of Stony Brook University in New York, ultimately suggested that “[deep learning (DL)]-based pathomics and classical image analysis appear useful in segmenting nucleated cells, fat cells, and evaluable BM regions and characterizing the cellular features of the BM tissue microenvironment.”

Researchers used morphology- and color-based models to differentiate trabecular bone, evaluable BM regions, and adipose tissue, which allowed them to estimate the ratio of hematopoietic cells to fat and derive the proportion of segmented cell nuclei within evaluable BM regions to estimate nucleated cell density as a proxy for cellularity.

Estimating BM Cellularity in MPNs with Pathomics-based Imaging Shows Potential

Whole slide images of BM biopsies with stained hematoxylin and eosin and respective clinical and pathological data were obtained from 54 patients with polycythemia vera (PV) at a university center in the United States. Among the 54 patients, 17% were newly diagnosed, 61% were pretreated, and 17% were relapsed.

Comparing three DL algorithms in addition to their tissue-based approach, the authors reported U-Net (r=0.7; ICC=0.6) and HistoCartography (r=0.7; ICC=0.5) algorithm estimations exhibited moderate agreement with hematopathology reports, while CellPose (r=0.2; ICC=0.2) and the tissue-based approach (r=0.2; ICC=0.0) had poorer agreement with BM reports.

Ultimately, the authors hoped to develop an image analysis strategy that expedited BM evaluation in order to eventually improve phenotyping of MPNs and other hematologic malignancies. “This pilot study demonstrates a practical and interpretable pathomics pipeline that serves as a proof-of- concept to support hypothesis-driven research in MPNs and other hematologic malignancies,” Krichevsky and colleagues concluded.

Reference

Krichevsky S, Abu-Zeinah G, Ouseph M, et al. Development of Pathomics image analysis tools to quantify bone marrow cellularity in myeloproliferative neoplasms. Abstract 138. 15^th International Congress on Myeloproliferative Neoplasms. November 2-3, 2023; Brooklyn, New York.