Abstract

Bone marrow reticulin fibrosis is associated with varied benign as well as malignant hematological conditions. The assessment of reticulin fibrosis is important in the diagnosis, prognostication and management of such disorders. The current methods for quantification of reticulin fibrosis are inefficient and prone to errors. Therefore, there is a need for automated tools for accurate and consistent quantification of reticulin. However, the lack of standardized datasets has hindered the development of such tools. In this study, we present a comprehensive dataset that comprises of 201 Bone Marrow Biopsy images for Reticulin (BoMBR) quantification. These images were meticulously annotated for semantic segmentation, with the focus on performing reticulin fiber quantification. This annotation was done by two trained hematopathologists who were aided by Deep Learning (DL) models and image processing techniques that generated a rough automated annotation for them to start with. This ensured precise delineation of the reticulin fibers alongside other cellular components such as bony trabeculae, fat, and cells. This is the first publicly available dataset in this domain with the aim to catalyze advancements the development of computational models for improved reticulin quantification. Further, we show that our annotated dataset can be used to train a DL model for a multi-class semantic segmentation task for robust reticulin fiber detection task (Mean Dice score: 0.92). We use these model outputs for the Marrow Fibrosis (MF) grade detection and obtained a Mean Weighted Average F1 score of 0.656 with our trained model. Our code for preprocessing the dataset is available at https://github.com/AI-in-Medicine-IIT-Ropar/BoMBR_dataset_preprocessing