Development of an automated artificial intelligence-based tool for reticulin fibrosis assessment in bone marrow biopsies.
Key claim: An InceptionV3-based fully convolutional network grades bone marrow reticulin fibrosis (MF0–MF3) on whole-slide images with excellent agreement to expert hematopathologists (Cohen’s κ=0.831).
Abstract
Bone marrow fibrosis plays a critical role in the diagnosis, prognosis, and management of haematological disorders, particularly myeloproliferative neoplasms like primary myelofibrosis. Accurate assessment of fibrosis, typically graded through histochemical techniques such as reticulin and trichrome staining, is essential but remains highly dependent on the pathologist’s experience. To address the challenges of variability in interpretation and the increasing demand for standardized evaluations, we developed a digital pathology system for automated bone marrow reticulin fibrosis grading. This study utilized 86 bone marrow biopsy specimens from patients diagnosed with Philadelphia chromosome-negative myeloproliferative neoplasms, collected between 2018 and 2023. A fully convolutional network based on the InceptionV3 architecture was trained to assess fibrosis grades (MF0-MF3) from whole slide images of reticulin-stained sections. The model was trained using 3814 annotated images and validated using a separate set of 40 BMBs. The algorithm’s performance was evaluated by comparing its fibrosis grading to expert hematopathologists’ assessments, yielding a Cohen’s kappa coefficient of 0.831, indicating excellent agreement. The algorithm showed strong concordance in fibrosis grading, especially for MF0 (k = 0.918) and MF3 (k = 0.886), and substantial agreement for intermediate grades (MF1 and MF2). Further validation across multiple institutions and scanning platforms confirmed the algorithm’s robustness, with an overall agreement of 0.816. These results demonstrate the potential of digital pathology tools to provide standardized, reproducible fibrosis grading, thereby aiding pathologists in clinical decision-making and training.
Why this matters
Reticulin fibrosis grade (MF0–MF3) is the histopathologic backbone that separates ET from preMF from overt PMF and feeds directly into risk models like MIPSS70+v2, yet inter-observer variability among hematopathologists has long limited its reproducibility. A validated AI tool with κ=0.831 concordance could standardize subtype assignment at the fibrosis boundary and enable more consistent longitudinal tracking of fibrosis progression or regression — the latter increasingly relevant as JAKi and non-JAK combinations (pelabresib, pacritinib, tasquinimod) claim fibrosis-modifying effects.