Computer Vision News - March 2022

31 Naronet: Discovery of Tumor Microenvironment In the previous figure, you can see the process of training NaroNet . It learns a mapping that relates patient information with patient labels. The architecture of NaroNet is divided into two consecutive networks G which is an ensemble of three parallel networks that assigns nodes to distinct P, N, A values. The second section, f2, assigns the patient's predictions from the learned TMEs. To learn the specific tumor microenvironment, the three neural networks were used with the previously provided data and later pooled to obtain the abundance of each TME. Phenotype learning, neighborhood learning (with a graph neural network, GNN) and area learning were utilized as extra elements in the NaroNet. NaroNet's classification accuracy (and 95% confidence interval) and interpretability were calculated as the intersection of the most relevant extracted TME and the ground truth of each synthetic experiment. Thenetworkparameters andarchitecture variations areoptimally selected by an architecture search algorithm. It identifies elements of the tumor landscape that relate to a specific predictive task. The patient's predictions are made solely using the relative abundance of TMEs. The model is evaluated with the entire patient cohort, and new prediction probabilities are obtained. If the null hypothesis is accepted, the extracted TME is considered to have a predictive value. In the following figure, you can see the association of high-grade endometrial carcinomas with patient-level labels. The ROC curves show the classification performance of NaroNet for the four tissue characteristics learned, the neighborhood composition of area A1 and patches assigned to a specific phenotype. Again, see more details of this in the original article!