Abstract 154: Improving MultiOmyxTM Analytics cell classification workflow efficiency by Invariant Information Clustering on historical data

• Published in: Cancer research (Chicago, Ill.) Ill.), 2021-07, Vol.81 (13_Supplement), p.154-154
• Main Authors: Reddy, Vivek, Stavrou, Nicholas, Nagy, Mate, Au, Qingyan
• Format: Article
• Language: English
• ISSN: 0008-5472; 1538-7445
• DOI: 10.1158/1538-7445.AM2021-154

Understanding the tumor composition and microenvironment provides insight into the efficacy and viability of immunotherapy treatments. Obtaining cell-level co-expression data (biomarker patterns expressed on cells indicating their phenotype) provides insight into where and how various cell types are distributed throughout the tumor. The MultiOmyx technology platform provides a detailed quantitative output including co-expression data through a multiplexed fluorescence microscopy approach, followed by imaging analysis using a proprietary analytics pipeline.The MultiOmyx analytics pipeline involves training a neural network for every biomarker of interest. Each biomarker's neural network will classify all cells as positive or negative for the marker. A major time-sink when constructing these neural network models is manually collecting and annotating training data. Moreover, new models must be built for the same biomarkers across different studies due to variability between samples. However, previously-built models can be applied on new images provided that their training image data and the new images follow the same distribution. This eliminates the bottleneck step of training a new model. In order to find ideal existing models for a new batch of samples, Invariant Information Clustering can be used to identify similarity of the new batch to previous images. Invariant Information Clustering (IIC) is used to group historical images by visual appearance. Image data is collected from the internal database, which stores all images from past studies. These images are split into image tiles (to avoid memory overload) and are then clustered with IIC. IIC works by generating image pairs that consist of the original image and an augmented version. It then maximizes the mutual information between the image and its transformed version using a neural network with a Resnet-34 architecture. This results in extracting the meaningful parts out of the original image while discarding the instance specific details. Augmentations including rotations, flipping, cropping, and noise are used to guide the network to learn proper features and be invariant to others. Visually similar tiles form distinct and tightly-bound cluster groups, each of which has a model assigned to it. When running inference, new image tiles are passed through the network and are assigned to the closest cluster. Cells from each tile are then classified using the model which corresponds to that tile cluster.