Spatial Mapping of Myeloma Bone Marrow Microenvironment Using a Deep Learning Approach

Background: Interactions between multiple myeloma (MM) cells and the bone marrow (BM) microenvironment influence disease evolution and drug resistance. Information is largely based on BM aspirates, which lack topological information and can be affected by haemodilution. To examine the spatial relationships between tumour and immune cells in the MM immune tumour microenvironment (iTME), we used multiplex immunohistochemistry (MIHC) coupled with deep learning (DL) image analysis for detailed visualization and unbiased automated analysis of BM components. Methods: BM trephines from patients (at a single centre) with monoclonal gammopathies of undetermined significance (n=9, MGUS) and newly diagnosed MM (n=10, NDMM) with paired baseline and day100 post autologous stem cell transplantation (ASCT) were stained with MIHC panels (BLIMP-1 for tumour, CD4, CD8, FOXP3). To dissect the unique BM landscape, we developed a superpixel based DL model (MoSaicNet) to distinguish blood, bone, fat and cellular tissue. We then developed a cell abundance aware DL method for cell detection and classification (AwareNet) on whole slide images. To tackle cell abundance bias, AwareNet assigned higher attention scores to rare cell types during model training. We used an independent validation cohort consisting of 9 NDMM with paired baseline and post-treatment BM samples to further evaluate the model's performance (n=18). Validation cohort samples were obtained from 7 different hospitals and were stained using a different autostainer of the same model. A colour normalization step was added before analysis. Results: MoSaicNet and AwareNet enabled us to evaluate densities and spatial proximity between cell types. To train, validate and test MoSaicNet, 269 regions of interest (69884 superpixels) from 19 BM samples were annotated by pathologists. For AwareNet, 8004 single-cell annotations were made on 11 samples. MoSaicNet achieved an overall classification area under the curve (AUC) of 0.99, 95% confidence interval (CI) [0.989, 0.991]. Similarly, AwareNet achieved a high overall AUC of 0.98 [0.977, 0.984]. AwareNet outperformed state of the art methods such as U-net and CONCORDe-Net in detecting both abundant and rare cell types. In addition, both algorithms achieved an overall AUC of >0.97 when applied to the validation cohort, despite different tissue processing and staining methods. Densities of FOXP3+CD4+, FOXP3-CD4+ and CD8+ cells did not differ significantly between NDMM and MGUS samples (p=0.32, p=0.81 and p=0.74 respectively), however, there were fewer BLIMP1+ cells in proximity to CD8+ cells in MGUS samples compared with NDMM samples. In NDMM patients, post-treatment samples revealed a reduction in densities of BLIMP1+ cells (p=0.013), CD8+ T cells (p=0.004) and FOXP3+CD4+ regulatory T cells (p=0.004) when compared with baseline. The number of BLIMP1+ cells in proximity to CD8+ and CD4+ cells significantly reduced post-treatment (Corrected for densities, p<0.05), indicating changes in iTME. Analysis of the validation cohort provided similar density and spatial results. Conclusion: We describe the use of a DL pipeline for tissue segmentation and automated cell annotation, enabling spatial mapping of the MM iTME to address topographical questions related to immune function and cell-to-cell interactions. The high accuracy in the validation cohort suggests that our model could be applied to other independent datasets after a normalization step.