An Efficient Data Representation for Histologic Images and its Applications in Biomedical Research

Abstract

Histologic imaging is a cornerstone of biomedical research, but the analysis of terabyte-sized data sets of high-resolution images is computationally challenging. A common solution relies on the identification of individual biological cells and the subsequent extraction of relevant features from this list of objects. However, this form of representing the image data is non-spatial in nature and therefore complicates the hypothesis-free analysis of local cell-neighborhoods – an important step for the study of tumor micro-environments and other applications. In this work, we propose a novel way of representing histologic image data by using the list of identified biological cells and turning it into a spatial image-format which uses a much lower resolution than original input images. These so-called Cell2Grid images are directly suited for training convolutional neural networks and offer similar evaluation metrics for traditional histologic image analysis pipelines. We applied our method to real-world examples by using two different image data sets and compared the results of image analysis pipelines with state-of-the-art methods. One core finding was that Cell2Grid images improved the predictive performance of a machine learning model compared to using conventional image rescaling methods. At the same time, it could reduce the computational cost of training such models compared to using high-resolution raw images, thus making training deep learning models feasible on off-the-shelf computer hardware. In addition, this work illustrates how the proposed concept of representing histologic images leads to a series of additional applications, including the procedural generation of synthetic image data and an efficient way to create three-dimensional tissue models from consecutive tissue sections. Finally, it is discussed how Cell2Grid data representation may in the future lead to a better understanding of biological processes by facilitating the systematic interpretation of trained neural networks on a cell-level to harness the full potential of histologic imaging.

Date of Award	2024
Original language	English
Awarding Institution	Medical University Graz
Supervisor	Pablo Lopez Garcia (Supervisor) & Thomas Pieber (Supervisor)