Computer Vision News - March 2022

30 Medical Imaging Research In this figure, you can see the graphical description of the CCI1 experiment. The ground truth on (a) define each patient type which after the processing described earlier leads to classification with the squares showing patches assigned to the learned neighborhood (N9) located in the GT neighborhood Nb2. The goal of the first step of the pipeline is to convert each high dimensional multiplex image of the cohort into a more manageable list of low-dimensional embedding vectors. To this end, each image is divided into patches - our basic units of representation of the local tissue microenvironment, or phenotype - and each patch is converted by the PCL module. The PCL module is trained iteratively. In each iteration, the PCL module is unsupervised trained to learn embeddings of a random set of patches. The choice of the image patch size S L is critical as it determines the extent to which biological structures can be captured. A multi-layer perceptron (MLP) maps each representation to a 128-dimensional vector to create similar embeddings for patches contained in the same crop. Agraph is then created that contains all the embedded patches of each tissue/image capturing cellular neighborhoods. This graph is G = (Z, A) where Z is a matrix that contains all the embeddings of the image. A is an adjacency matrix that contains the connectivity between patches.