41 Cracking the Code Computer Vision News Computer Vision News within the tumor microenvironment and their notorious heterogeneity. Traditional diagnostic methods involve examining slides under a microscope, which is time-consuming and prone to human error. Even seasoned pathologists can disagree on the same samples, highlighting the need for precise computer-aided diagnosis techniques. The efforts focused on developing deep learning models that can automatically grade gliomas by analyzing tissue microarrays with human leukocyte antigen staining. Recent results presented at SPIE Medical Imagingand published in the Journal of Imaging Informatics in Medicine investigate the strategies for glioma multiclass classification - both in supervised and weakly supervised manner with single-cell analysis. One challenge in this research was the limited data available for training the models. The issue was tackled by augmenting underrepresented classes and evaluating models using k-fold stratified cross-validation. The DenseNet121 architecture with prior image preprocessing outperformed the baseline model increasing accuracy by 9%. The study goes beyond identifying the WHO grade (1, 2, 3, 4, or grade “0”) to which the sample should be assigned. The protocol of weakly supervised deep learning with singlecell analysis was applied to the dataset to learn, discover, and quantify the cell phenotypic neighborhoods across grades. The model identified two critical cell clusters, called neighborhood N2 and N4, which showed significant differences in the abundance of cells, most likely microglia and brain Graphs indicating significant differences in the abundance of given cell clusters across grades.
RkJQdWJsaXNoZXIy NTc3NzU=