The Philadelphia-negative myeloproliferative neoplasms (MPN), including polycythemia vera (PV), essential thrombocythemia (ET), and primary myelofibrosis (MF), are clonal hematopoietic stem cell disorders characterized by excessive proliferation of myeloid lineages. Splanchnic vein thrombosis (SVT) is a rare site of thrombosis frequently associated with MPNs (MPN-SVT) 1. Sub-classification of MPN in this setting can be challenging because of significant phenotypic overlap, blood counts not meeting diagnostic thresholds, and atypical bone marrow (BM) features 2, 3, resulting in frequent designation as MPN-unclassifiable (MPN-U) 4, 5. Histopathology remains core to the diagnosis of MPN; however, it is limited by interobserver variability 6-8. Artificial intelligence (AI) approaches can augment pathological evaluation by objectively identifying and quantifying key diagnostic features such as megakaryocyte cytomorphology/topology and fibrosis severity. They have also shown potential to provide additional prognostic information 9-11. We evaluated diagnostic concordance between clinical, conventional histology, and AI-powered BM evaluation in patients with MPN-SVT 9, 12. Patients diagnosed with MPN-SVT attending the Royal Free Hospital NHS Trust between 2000 and 2015 were included, subject to the availability of clinical, laboratory data, and BM trephine biopsies. Data at the time of SVT, MPN diagnoses, and BM assessment were collected retrospectively. Clinical diagnosis was based on clinical features, peripheral blood parameters, and BM findings (WHO 2008 criteria). This was assigned around 3–6 months after MPN diagnosis, when clinical status and blood counts stabilized. At study inclusion, cases were also assigned a WHO 2016 diagnosis. Archived BM trephine biopsies were independently reviewed by an expert hematopathologist blinded to clinical and laboratory data (except age to help assess marrow cellularity), and a favored histological diagnosis was assigned 12. Anonymized digitized images of BM slides were analyzed by AI, assessing megakaryocyte and fibrosis features. A favored AI diagnostic label using feature similarity to an external reference Oxford MPN cohort was proffered by manually combining the AI-based outputs of megakaryocyte and fibrosis algorithms (Figure S2) 9, 10. To demonstrate the applicability in the local cohort, these algorithms were also applied to a separate validation cohort of 30 patients with MPN without SVT (Table S3). The study is descriptive, expressed as a range for continuous variables and as percentages for categorical variables. JAK2 allele frequency was compared using the two-tailed Mann–Whitney U test, with 𝑝 ≤ 0.05 considered significant. Thirty-four patients with MPN-SVT were included, with four patients undergoing two BM assessments. Two samples were unsuitable for analysis, yielding 36 cases (Tables S1 and S2). Median follow-up was 13.2 years (5.7–33.6 years). Median time to BM assessment was 3 months post-MPN diagnosis, with 58% and 70.9% performed within 6 months and 1 year, respectively. Seven patients were on cytoreductive treatment at the time of sampling. Clinical diagnosis was considered the benchmark as it determined treatment decisions. Clinically, cases were diagnosed as PV (15 cases), ET (8), MF (9), and JAK2V617F with normal counts (4). Using the WHO 2016 criteria, MPN-U was the predominant diagnosis (12 out of 36 cases). Blinded expert histopathology classified cases as PV (14 cases), ET (4), prefibrotic-MF (6), MF (6), MPN (4), reactive (1), and MDS/MPN (1). Clinical diagnosis had higher concordance with expert histology (50%) than with the original BM reports (35.3%). Based on the combined principal component analysis (PCA) distribution AI-analysis of megakaryocyte and fibrosis features (relative to the centroids of disease space), 12 samples clustered near the PV region, 16 near the prefibrotic-MF region, 5 near the MF region, and 3 near the reactive region, with none near the ET region (Figure 1). Concordance across all four modalities was low at 33.3% in MF and PV; none in ET and JAK2V617F with normal counts. Histology achieved 66.6% concordance with clinical PV and MF, but only 25% with clinical ET. The combined PCA-based AI feature cluster positions corresponded with clinical diagnoses in 46.6% for PV, 33.3% for MF, and none for ET. Overall, AI feature-space positioning agreed with histology in 41.6% (15 out of 36) of cases, approaching 50% in PV. Figure 2 reveals two patterns (especially in PV) wherein a subset of patients has a consistent diagnostic label, and another subset shows discordance across the four methodologies. ET: Clinical ET did not concord with AI analysis and histology categorization. Histology identified only 2 of 8 clinical cases as ET; the remaining 6 were reclassified as prefibrotic-MF (3), PV (2), and MPN (1). On AI analysis, again, none were positioned within the ET region; instead, they clustered closer to the prefibrotic-MF (5) and PV regions (3). Notably, AI assessment of megakaryocyte features was associated with non-ET characteristics, with significant overlap into prefibrotic-MF (Figure 1), suggesting a subtle difference in the histopathological process compared to conventional ET. This observation is significant, given that the AI findings showed 50% concordance in ET from our validation cohort (For analysis of the validation MPN cohort, refer to the Supporting Information file: Table S3 and Figure S1). It also raises the question about the nature of thrombocytosis in MPN-SVT. Inflammation and iron deficiency could contribute to relatively low hematocrit and high platelets. Prefibrotic-MF is also an important cause of thrombocytosis, which would be consistent with the AI interpretation. PV was the most frequent clinical diagnosis (15 of 36 cases, 42%), with concordance on histology in 10 cases. On AI analysis, seven cases of clinical PV showed concordant positioning in the combined AI-derived feature space; however, the remaining were labeled as prefibrotic-MF (5), MF (1), and reactive (2). This suggests that there is heterogeneity in clinically diagnosed PV cases on both histology and AI analysis. Prefibrotic-MF was the most prominent category by AI analysis (16 cases), corresponding to 5 cases each of clinical PV, ET, MF, and 1 case of JAK2V617F with normal counts. In comparison, histology identified only 6 cases as prefibrotic-MF, WHO 2016 diagnosis of 2 cases, and no cases in clinical diagnosis. MF comprised nine cases in clinical diagnosis. In the PCA analysis of AI-derived fibrosis features (Figure 1 and Figure S2), only three cases were positioned within the MF region, while five were positioned in the prefibrotic-MF region, suggesting a lower fibrosis severity. Eight of nine cases were positioned within the MF region based on megakaryocyte features, but composite analysis favored prefibrotic-MF in five cases. In contrast, the expert histology identified MF in six cases. The discordance between AI-based and histological categorizations in prefibrotic-MF likely arises from the difficulty in manually distinguishing between fibrosis grades 1/2 and identifying the subtle megakaryocyte features of prefibrotic-MF. Notably, samples positioned within the prefibrotic-MF and MF regions of AI feature space showed a higher median JAK2 VAF (35%) compared with the rest of the cohort (6%) (p = 0.02). These findings raise the possibility that prefibrotic-MF features in this cohort could indicate either a progression from PV/ET or represent a primary disease process. One mechanism underlying this could be the endothelial damage, which predisposes to both fibrosis and splanchnic thrombosis 13. In a registry study of MPN-SVT, the survival was shortened due to hepatic disease and bleeding rather than the evolution of MPN 14, suggesting that fibrosis may represent a distinct process rather than advanced disease. MPN-U comprised 33% (12 cases) of the WHO 2016 diagnoses. AI-derived feature analysis showed heterogeneity, PV (5), prefibrotic-MF/MF (4), and reactive (3) (Figure 1). The diverse reassignment using AI-based assessment is noteworthy, though such cases were excluded in the validation of the original model 9. These features hint at characteristics that may be intrinsic to this subset and might explain the better outcome reported in MPN-U 15. For further analysis, refer to the Supporting Information. All four cases with JAK2V617F with normal counts had abnormal BM findings by histology and/or AI, indicating the presence of a latent phase MPN rather than JAK2 CHIP. Sequential BM assessments in four patients demonstrated the utility of AI analysis in evaluating disease progression, particularly worsening fibrosis, for example, cases 17 and 24, with progression of fibrosis in case 24 (Figure S3). Our study provides insights into the interpretation of BM histology in a well-characterized cohort of MPN-SVT. The AI model identifies distinct and subtle features that distinguish this cohort from classical MPNs. Among the diagnostic reassignments, the paucity of ET on expert histology review and its complete absence by AI assessment are striking. This is further supported by the identification of ET-like features within the validation cohort of annotated patients with ET. In contrast, prefibrotic-MF-like features are enriched in the MPN-SVT cohort. These features span patients classified clinically as MF, PV, and ET. Quantitative evaluation of megakaryocyte and fibrosis features has the potential to reduce inter-observer variability. AI-based feature analysis can visualize the spatial localization of fibrosis and temporal changes over time using heat maps, suggesting a complementary interpretive role for AI in follow-up. The limitations of the study include that AI models do not incorporate other MPN features, such as panmyelosis typical of PV or granulocytic proliferation of prefibrotic MF, retrospective design, small sample size, and variability in the timing of BM sampling. Nonetheless, this study demonstrates that AI-based analysis can complement histology, and larger prospective studies are warranted to validate these findings. The authors have nothing to report. This project was conducted as part of a departmental service improvement initiative, and all patient data were anonymized at the source in accordance with the United Kingdom Data Protection Act (1998). Declaration of interest for Daniel Royston: D.R. provides consulting services to Ground Truth Labs Ltd and Johnson and Johnson, and is co-founder of Ground Truth Labs Ltd. The data that support the findings of this study are available from the corresponding author upon reasonable request. Table S1: Clinical and laboratory characteristics of MPN-SVT patients. Table S2: Clinical diagnosis and molecular markers in the MPN-SVT cohort. Table S3: Clinical diagnosis and molecular markers in the MPN cohort. Figure S1: Summary of MPN (Validation cohort) diagnostic categories for each sample based upon clinical (C), histology only (H), WHO 2016 diagnosis (WHO 2016), AI consolidated (AI—consol), AI fibrosis, and AI megakaryocytes (AI—Megak) of MPN. Clinical diagnosis (C) refers to the pragmatic clinical diagnosis assigned within 6 months of MPN diagnosis by the treating clinician for patients. The WHO 2016 diagnosis was assigned retrospectively by an independent reviewer (BS) at the time of study inclusion, based on the original BM reports. Histopathology (H) comprises a review of the bone marrow trephine slides by an independent, blinded histopathologist. AI-fibrosis and AI-megakaryocyte denote the artificial intelligence interpretation of the fibrosis and megakaryocyte morphology on the trephine slides. AI-consolidated represents a pragmatic, manual integration of the AI-fibrosis and AI-megakaryocyte analyses, as outlined in the main text. Figure S2: Principal component analysis (PCA) representations of AI analysis of MPN-SVT samples. Top (a): Megakaryocyte feature analysis combining cytomorphological features and topological distribution. Bottom (b): Fibrosis feature analysis combining average tile continuous indexing of fibrosis (CIF) score, tile score distribution, and heterogeneity of CIF score. To simplify this output and provide a favored or estimated AI-based categorization for each case, the nearest Euclidean distance between the sample features and the center of each MPN subtype distribution in PCA space for both megakaryocyte and fibrosis features. Figure S3: Visual representation of paired bone marrow (BM) images of a patient (left) and principal component analysis (PCA) of AI fibrosis, AI megakaryocyte, and AI composite analysis (right), labeled case 17 (above) and case 24 (below). Cases 17 and 24 pertain to a patient who underwent marrow examination 6 years apart (in 2011 and 2017) with a diagnosis of CALR-mutated ET on clinical and histological criteria but labeled as prefibrotic-MF on AI. AI fibrosis shows progression in the intervening years. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.
Sreedhar et al. (Thu,) studied this question.