Computational Cytomorphological Analyses Identifies Bone Marrow Erythroblast Proportion As a Biomarker of Treatment-Free Remission in CML

Introduction

Tyrosine kinase inhibitor (TKI) treatment can be safely discontinued in chronic-phase chronic myeloid leukemia (CP-CML) patients who achieve deep molecular response. Several clinical trials have demonstrated persisting treatment-free remission (TFR) rates of ~50% following discontinuation, consolidating this as a novel goal of therapy. The duration and depth of response prior to discontinuation, mutations in epigenetic modifier genes, NK cell phenotype, and BCR::ABL halving time have been suggested as biomarkers illustrating the complexity of mechanisms regulating TFR. Cytomorphology slides represent a globally accessible and affordable resource, but their utility to predict TFR has not been explored due to the lack of high-resolution imaging data.

Methods

We included 110 CP-CML patients having attempted first TKI discontinuation in routine practice or in clinical trials from 6 university hospitals (Adelaide, Australia n=63; Saga, Japan n=12; Ome, Japan n=12; Trondheim, Norway n=9; Uppsala, Sweden n=8; Örebro, Sweden n=6). We digitized diagnostic cytomorphological bone marrow (BM) slides at 10x magnification and representative areas separately with 100x magnification. To comprehensively analyze the BM cytomorphology, we analyzed images with the Hematoscope application (unpublished) operating 15 deep learning-based image analysis algorithms, performing BM sample segmentation, detection and classification of 17 cell types, and dysplasia detection. Due to poor sample quality, 8 samples were excluded from analyses.

To identify novel biomarkers of TFR, we compared the distribution of cytomorphological features in CML patients relapsing from vs. maintaining TFR, with a minimum follow-up time of 12 months (Wilcoxon signed-rank test). We also integrated TKI treatment (generation, depth, duration), response (depth, duration), and CML risk scores and their components (ELTS, Sokal, Hashford, EUTOS) to compare the distribution of significant cytomorphological features to established clinical features.

Results

The proportions of erythroblasts and granulopoietic cells were the cell types most significantly linked to TFR. Relapsed patients (n=51) had a higher proportion of erythroblasts (median 7.8%, 25-75% range [5.4 – 11.2%]) compared to patients maintaining TFR (n=51; 5.8% [3.6 – 8.4%], p=0.013). In contrast, relapsed patients had a lower proportion of granulopoietic cells (87.1% [82.7 – 89.9%] vs. 89.5% [85.2 – 91.7%], p=0.070). The contrast was clearer when comparing patients with early relapse ≤6 months (n=38) to patients maintaining TFR (n=51). Patients in the early relapse group had a higher proportion of erythroblasts (8.4% vs. 5.8%, p=0.005) and a lower proportion of granulopoietic cells (87.0% vs. 89.5%, p=0.055). Dysplastic erythroblasts (e.g., megaloblastic, multinucleated, vacuolization) and megakaryocytes (e.g. hypolobulated, separated nuclei) were uncommon, and not associated with relapse.

Given the absence of high-risk ELTS patients in TKI discontinuation trials, we then compared the proportion of these cell types in low (n=73) vs. intermediate-high risk groups (n=29). We observed an enrichment of erythroblasts (7.7% vs. 5.3%, p=0.004), and lower proportion of granulopoietic cells in low-risk patients (87.2% vs. 90.4%, p=0.001). Longer duration of MR4.0 before TKI discontinuation was not found to be significant (40.1 vs. 47.6 months, p=0.12).

Finally, we examined the association of BM cells to TFR maintenance by the last-line TKI treatment before discontinuation. In total, 57 patients were treated with imatinib and 45 patients with a second generation TKI (dasatinib, nilotinib, bosutinib). Relapsed imatinib-treated patients had a higher proportion of erythroblasts (7.8% vs. 4.5%, p=0.047), and a lower proportion of granulopoietic cells (87.2% vs. 89.7%, p=0.052). No association was found in the patient group treated with a second generation last-line TKI.

Conclusion

The results suggest that hematopoietic lineage distribution and maturation at diagnosis could influence TFR maintenance and possibly guide treatment decisions in CP-CML patients. We demonstrate the potential of BM cytomorphology and modern computational techniques to identify novel clinical biomarkers. Future work will aim to further validate the results in additional datasets, and train a multivariate classification model to predict TFR.