Overview of computational deep learning and image processing pipelines for BM MIHC images: A, MoSaicNet pipeline. The polygons (black) indicate superpixels. MoSaicNet dissects a tissue section into bone, blood, fat, and cellular tissue regions (Supplementary Materials and Methods). B, AwareNet for attention-based cell detection and classification (Supplementary Materials and Methods). The attention image pixel values were generated from the abundance of cell types. An attention image was applied to the objective function during model parameter optimization to regularize the algorithm by assigning high attention to rare cell types. The cell detection algorithm generates a cell probability map. A postprocessing algorithm was developed to find the cell nucleus center, (x, y) location, from the probability map (Supplementary Materials and Methods). A patch centered on each cell was extracted and fed to deep learning (DL)-based classifier to infer its class. C, Spatial and morphologic analysis of BM trephine samples. Bone texture and structural heterogeneity were investigated using an autoencoder-based machine learning method (Supplementary Materials and Methods). We used spatial proximity analysis to study the spatial relations of cells. r, radius. Cell density refers to the number of cells per unit of tissue area.
ARTICLE ABSTRACT
Bone marrow trephine biopsy is crucial for the diagnosis of multiple myeloma. However, the complexity of bone marrow cellular, morphologic, and spatial architecture preserved in trephine samples hinders comprehensive evaluation. To dissect the diverse cellular communities and mosaic tissue habitats, we developed a superpixel-inspired deep learning method (MoSaicNet) that adapts to complex tissue architectures and a cell imbalance aware deep learning pipeline (AwareNet) to enable accurate detection and classification of rare cell types in multiplex immunohistochemistry images. MoSaicNet and AwareNet achieved an AUC of >0.98 for tissue and cellular classification on separate test datasets. Application of MoSaicNet and AwareNet enabled investigation of bone heterogeneity and thickness as well as spatial histology analysis of bone marrow trephine samples from monoclonal gammopathies of undetermined significance (MGUS) and from paired newly diagnosed and posttreatment multiple myeloma. The most significant difference between MGUS and newly diagnosed multiple myeloma (NDMM) samples was not related to cell density but to spatial heterogeneity, with reduced spatial proximity of BLIMP1+ tumor cells to CD8+ cells in MGUS compared with NDMM samples. Following treatment of patients with multiple myeloma, there was a reduction in the density of BLIMP1+ tumor cells, effector CD8+ T cells, and regulatory T cells, indicative of an altered immune microenvironment. Finally, bone heterogeneity decreased following treatment of patients with multiple myeloma. In summary, deep learning–based spatial mapping of bone marrow trephine biopsies can provide insights into the cellular topography of the myeloma marrow microenvironment and complement aspirate-based techniques.
Spatial analysis of bone marrow trephine biopsies using histology, deep learning, and tailored algorithms reveals the bone marrow architectural heterogeneity and evolution during myeloma progression and treatment.