Computer Vision News - March 2022

32 Medical Imaging Research More synthetic experiments were performed on the network, and it was measured on endometrial carcinomas. The network unsupervised learned 26 TMEs. The area A1 (p-value: 2.56x10 -9 ) seemed the most predictive TME when making patient predictions. A1 is significantly more associated with tumors harboring POLE mutations. To illustrate the individual interpretability of the results, show two examples of images in which phenotype P2 was the most relevant TME selected by NaroNet. Summarizing NaroNet is the first WSDL method fully adapted to multiplex imaging. Two state-of-the-art WSDLmethodswere used to classifyH&E tissue sections, adapted for the analysis ofmultiple images. ClAM is basedon a two-step strategy. In the first step, the image is divided into image patches (i.e., hundreds of cells) which are fed to a ResNet50 pre-trained on ImageNet. In the second step, attention scores are assigned to patch representations. NaroNet as an end- to-end deep learning framework proves this hypothesis true. It accurately performs patient predictions from local phenotypes, neighborhoods, and areas. One of themajor bottlenecks in developing high-performance machine learning classifiers for computational pathology is