Computer Vision News - March 2022

29 Naronet: Discovery of Tumor Microenvironment But youwill askwhat exactly is this paper trying to do? NaroNet is a multilevel, interpretable deep learning ensemble . It learns the most relevant TMEs from multiplex immunostained tissue sections . It is a computer model that predicts TMEs and predicts clinically relevant parameters and is based on synthetic sets of multiplex images. Methods and approaches NaroNet integrates novel and state-of-the-art ML approaches. The development of patch contrastive learning (PCL), a self-supervised learning algorithm that encodes high- dimensional pixel information into enriched patch-embeddings. The paper explores the following important elements and is structured as follows: first, the synthetic and real datasets used are explained, followed by the proposed methodology. The next section contains the experiments used to test the performance of NaroNet and reports the results obtained. The following section provides an in-depth analysis of the proposed methods. Finally, the results, conclusions and future ideas are explored. Let’s see a few of them! First, let’s talk about the data. Synthetic patient cohorts . An in-house developed multiplex immunostained tissue simulator was used to create patient cohorts. Each patient of the cohort was represented by an 800x800 multiplex image. 7 patient cohorts were simulated. Each cohort contained 240 patients, distributed in 3 groups (type I, II, and III) of 80 patients eachwith different disease paradigms inspired by real scenarios. Phenotype Marker Intensity (PMI) was used to determine the relative intensity of Mk6 marker expression in each group of patients. Four neighborhoods are defined based on the relative abundance of the phenotypes. Nb3 was set to 15% (moderately present), whereas in PMI2 the relative abundance of Ph6 was set to 0.25% (rarely present). Phenotype Frequency (PF) was set to 0% (type I), 30% (type II), and 60% (type III) % and in relevant ratios for the other frequencies (you can read in more detail on the paper) Neighborhood-Neighborhood Interactions (NNI) was designed to simulate different interactions between cellular neighborhoods, related to patient type. Nb2 and Nb3 repel (type I), show no interaction (type II), or attract (type III). Next, tissue sections from twelve high-grade endometrial carcinomas were stained with a seven-color multiplex panel targeting key elements of the immune environment and 336 with a size of 1876x1404x7 pixel images were obtained. The image dataset is publicly available, and the tissue sections were stained with a 35-plex antibody panel. Let’s see an image describing the experiment!