Gigapixel Patch Semantic Segmentation for Histopathology
Event Type
Research Poster
TimeTuesday, June 23rd3:35pm - 3:40pm
LocationAnalog 1
DescriptionWe study the impact of various-sized input patches and sub-sampling settings, when performing pixel-level semantic segmentation onhematoxylin and eosin (H&E) stained whole-slide images of lymph nodesections from the CAMELYON16 and CAMELYON17 dataset. Our re-sults show an increase in final mean intersection over union values by25% when going from 256x256 pixel patches to larger 2048x2048 pixelones. What is remarkable for this type of data is that images are cap-tured at a gigapixel resolution. Our results show that working with these2048x2048 patches leads to a memory footprint of approximately 25GBof memory per example for our extension of the scale invariant deeplab-V3+ convolutional neural network architecture. Finally, we also presentour scale-out results up to 256 nodes enabled by Intel Xeon Scalable Per-formance multiprocessors. Finally, we define as goal performing regionof interest – wise whole slide analysis with a maximum level of detailat gigapixel resolution, contributing to computer aided diagnosis for thecommon pathological practice.
Poster PDF
Poster Authors
Machine Learning Consultant
Machine Learning Consultant
Machine Learning Consultant
PhD candidate