Gliomas, aggressive, highly heterogeneous brain tumors, require precise grading for effective treatment. The role of myeloid cells in the tumor microenvironment (TME) is crucial for glioma progression and patient prognosis, emphasizing the need for advanced diagnostic tools.
Current glioma grading relies on manual histological evaluations, which are time-consuming and subjective, often missing intricate TME characteristics. Existing methodologies using deep learning models, such as convolutional neural networks and vision transformers, face challenges in multiclass grading due to the need for large annotated datasets and their limited ability to capture complex TME.
Our study in collaboration with Nencki Institute of Experimental Biology proposes using single-cell analysis with unsupervised deep learning to analyze glioma tissue microarrays stained for Human Leukocyte Antigens, focusing on myeloid cell accumulation. This method differentiates glioma grades by identifying unique TME phenotypic neighborhoods. Despite a small dataset and the challenge of distinguishing between WHO grades 2 and 3, our analysis successfully identified distinct phenotypic neighborhoods, particularly N2 and N4, which are significant in differentiating malignant gliomas.
This approach demonstrates the potential of deep learning in accurately classifying gliomas and highlights the importance of myeloid cells in tumor progression. Our findings suggest that automatic grading based on phenotypic neighborhoods could significantly enhance intraoperative assessments and immunotherapy planning.
Our next step of the scientific investigation includes elucidating the mechanisms of immune evasion and resistance to immunotherapy in gliomas using computational methods with spatial transcriptomic data.
Additionally, in our computer vision studies for brain diagnostics, we propose a multimodal translation technique to generate brain histology from MRI, potentially avoiding invasive biopsy procedures. So far we tested generating synthetic histology of corpus callosum from MRI images.