The analysis and segmentation of digital pathology images has many challenges. Files are large and often contain imagery stored in a pyramid of multiple resolution levels. The most detailed of these routinely require 30 gigabytes of RAM or more to load fully into main memory and many times more to process naively. The multiple resolution levels do not always cover the same “real world” spatial extents and computing corresponding locations within these image levels is error-prone. Reading from some of these proprietary file formats can be difficult and slow.

Despite these hurdles, there are several tasks that researchers and clinicians routinely need to perform: segmenting objects, identifying features, aggregating details, labeling ground truth, and so on. Some of the most common end goals are assessing tissue health, evaluating drug efficacy, improving microscope design, etc. Many of these tasks are being automated with deep convolutional neural networks.

This talk will present some techniques to overcome the challenges listed above, with an emphasis on speed and throughput. It will also explore how the unique multi-resolution nature of pathology imagery can be used to improve results and reduce processing time. For example, identifying high information content areas in lower resolution views can limit time-intensive processing to only those regions. Additionally, we will show how to chunk up a single high-resolution image to create a large collection of normal sized images to train a semantic segmentation or object detection network… creating tens of thousands of training images even before data augmentation. We will pay special attention to the challenges of class imbalance.
Video created 01/19/2022

source