Correlating Image-Level Nucleomorphologic Features with Molecular Subtypes of Myeloproliferative Neoplasms

Introduction:

This study focuses on utilizing deep learning (DL) Pathomics image analysis tools to extract nucleomorphologic features from digital hematoxylin & eosin (H&E) whole slide images (WSIs) of histologic bone marrow (BM) core biopsies of Philadelphia chromosome-negative myeloproliferative neoplasms (Ph- MPNs). Nucleomorphologic features were clustered and aggregated to reveal correlations with molecular subtypes of Ph- MPNs. This work demonstrates a novel Pathomics workflow that can enhance subclassification by combining nucleomorphologic features with molecular testing to further our understanding of the biological behavior of Ph- MPNs.

Methods:

Following institutional review from Stony Brook University and Weill Cornell Medicine, we identified 271 marrow slides using the WMC MPN research data repository. These slides were from newly diagnosed patients with polycythemia vera (PV; n=83, 31%), essential thrombocytosis (ET; n=83, 31%), myelofibrosis (MF; n=76, 28%), and tumor-negative, normal controls collected for Non-Hodgkin's lymphoma staging (CTL; n=35, 13%). The clinical dataset included H&E-stained WSIs, laboratory testing, molecular studies, and pathology report elements.

Our previously reported Pathomics workflow utilizes a suite of tools to segment bony trabeculae and adipose tissue to identify red hematopoietic marrow in the BM microenvironment, followed by nuclear segmentation to compute BM cellularity by using nucleated cell density (NCD) and extract nucleomorphologic features [1]. Thirty million segmented nuclei were extracted from 271 WSIs and 106 nucleomorphologic features were analyzed and clustered. The nucleomorphologic features quantitatively describe nuclear color, size, shape, and texture. K-means clustering identified distinct groupings in two separate approaches. In the first approach, all nuclei from all cases were aggregated to examine whether nucleomorphologic features could distinguish different populations that might correspond with known attributes of erythroid, myeloid, lymphoid, and megakaryocytic cells. In the second approach, nucleomorphologic features for each image were clustered across cases to identify correlations with MPN molecular subtypes and investigate effects on overall survival.

Results:

Nuclear-level clustering revealed five unique clusters by uniform manifold approximation and projection (UMAP), suggesting partial separation between populations of nucleated cells. For further interpretability, clustering increased nuclear area and decreased nuclear circularity appeared important to distinguish cell populations by nucleomorphologic features, which were then mapped to the BM microenvironment to examine different clusters within a spatial context that is novel and still unexplored. Image-level clustering identified four unique clusters that correlate with Ph- MPN molecular subtype:

1. Cluster 1 (n=79) includes CTL and ET and MF cases that are JAK2-/CALR+
2. Cluster 2 (n=48) includes ET and MF cases that are JAK2+/TET2+
3. Cluster 3 (n=124) includes PV, ET, and MF cases that are JAK2+ only
4. Cluster 4 (n=22) includes MF cases that are JAK2+/ASXL1+

Subsequent analysis shows decreased overall survival in Cluster 2 (JAK2+/TET2+) and Cluster 4 (JAK2+/ASXL1+), which are implicated in progressive disease [2].

Conclusion:

This study presents a novel Pathomics image analysis workflow to correlate nucleomorphologic clusters with clinical outcomes in a scalable and interpretable manner, demonstrating the potential value of Pathomics in precision oncology for hematologic diseases. By integrating Pathomics with clinical informatics in this translational approach, future work will focus on elucidating the nuances of Ph- MPN pathobiology and disease progression with additional efforts in developing additional Pathomics tools and validation of additional independent datasets.

References:

[1] Krichevsky S et al. A deep learning-based pathomics methodology for quantifying and characterizing nucleated cells in the bone marrow microenvironment. Blood. 2023;142:2294.

[2] Grinfeld J et al. Classification and personalized prognosis in myeloproliferative neoplasms. N Engl J Med. 2019;379:1416-1430.