Reticulin-Free Quantitative Assessment of Bone Marrow Fibrosis in Myeloproliferative Neoplasms; Time to Sell the Family Silver?

Estimation of bone marrow fibrosis is central to the evaluation of patients with myeloproliferative neoplasms (MPN). However, European consensus criteria for fibrosis evaluation (MF grades 0-3) are subjective and fail to capture heterogeneity across fibrosis grades. We previously demonstrated the potential of machine learning to improve the detection and quantitation of marrow fibrosis using routinely prepared bone marrow trephine (BMT) samples [Ryou H, Leukaemia 2022; Ryou H, Hemasphere 2024].

Evaluation of marrow fibrosis requires standardized silver impregnation techniques that highlight the reticulin (type III collagen) component of the normal marrow extracellular matrix. These techniques can introduce artefacts such as silver droplet precipitates and bone peeling / tissue loss. In addition, variations in laboratory practice may under or overstain reticulin fibers and impart a range of colors to tissue sections. Such variation in reticulin stain quality is well recognised and generally exceeds the variation observed for Haematoxylin and Eosin (H+E)-stained sections in most diagnostic laboratories. However, this has not been systematically evaluated in the bone marrow.

In response, we have developed a machine learning (ML) model to generate a quantitative fibrosis prediction output directly from H+E-stained whole slide images (WSI). This model was trained and validated on 59 BMT samples from patients with either a reactive marrow (n = 6) or an established diagnosis of MPN (n = 13 essential thrombocytopenia [ET]; n = 15 polycythemia vera [PV]; n= 25 myelofibrosis [n = 19 MF & n = 6 pre-PMF]). Each sample was H+E stained followed by a destain step with subsequent staining for reticulin (Gordon and Sweet's method). This bespoke stain-destain-stain method enabled direct training of the H+E model from tissue used to generate our previously described reticulin-based fibrosis prediction (Continuous Indexing of Fibrosis; CIF), thus bypassing the need for an inferential training step. Image tiles (512 X 512 [113µm X 113µm]) were extracted from reticulin images, with CIF scores predicted for each tile as previously outlined [Ryou H, Leukaemia 2022]. An affine transformation matrix, derived from the segmented tissue masks of reticulin and H+E images, was applied to align reticulin tissue coordinates with corresponding H+E images to acquire matched H+E / reticulin tiles. The dataset was divided into a training set (30 samples: 9252 pairs) and validation set (9 samples: 1911 pairs) for H+E-based CIF model generation. A RankNet model was used to estimate fibrosis severity from the H+E-stained images, with a model prediction accuracy of 0.931. The average CIF score difference between the reticulin and H+E model prediction was small (0.061 ± 0.049). Importantly, the CIF maps generated from paired reticulin and H+E-derived images appeared similar, with detection of fibrotic foci across the marrow microenvironment.

Next, we sought to systematically compare the stain intensity variance between reticulin and H+E-stained BMT sections using 91 diagnostic samples (n= 37 ET, n = 18 PV, n = 21 MF, n = 3 pre-PMF and n = 12 reactive). Image tiles were extracted from WSIs, excluding tiles with bone coverage > 50%. Standard deviation of averaged intensity per RGB channel of each tile (excluding bone and fat) was used to measure intensity variance. The intensity variance of dominant colour was determined using K-means clustering to cluster the pixel intensities into 5 groups. Stain intensity analysis revealed higher variance in reticulin compared to H+E for all tile colors (R: 28.54 vs 16.28, G: 30.35 vs 25.64, B: 31.80 vs 16.06) and dominant colors (R: 27.79 vs 15.84, G: 31.47 vs 27.25, B: 33.55 vs 15.72).

In summary, we have developed an H+E-based model for fibrosis prediction designed to overcome some of the technical limitations of reticulin staining in BMTs, including variation in stain intensity and colour dominance. Our H+E-based CIF output demonstrated comparable performance to that of our previously published reticulin-based model across normal marrow and MPN subtypes. This model has the potential to significantly improve the speed and efficiency of quantitative marrow fibrosis assessment in MPN and improve QC performance within and between laboratories. It can also support advanced single-section imaging techniques designed to correlate morphological and gene-expression data with underlying fibrosis.