Sean Yu

M.S.

aetherAI

Data Scientist

NLST (Learn more about this study)

NLST-620

Dec 13, 2019

Training Lung Cancer Classifiers with Ultra-high Resolution Whole Slide Histopathology Images.

Analysis of digital whole slide images (WSIs) is difficult because of its extremely high spatial resolution, up to billions of pixels. Applying CNN to learn patterns on such high-resolution images is thus a challenging task. Most approaches require an inefficient pre-processing procedure that crop a WSI into tens of thousands of small patches (normally 256 × 256) and annotate them for each WSI beforehand. These patch-based methods have yielded some successful results. However, the ground truth for each image patch needs to be given, which is typically done by free-hand contouring on the whole slide images. This annotation process is extremely laborious. Furthermore, borders between different tissue classes are often difficult to identify, leading to inconsistent annotation between pathologists. Lastly, the high variability of tissue morphology makes it difficult to cover all possible examples during annotation and to sample representative patches during training.
To deal with these drawbacks, we utilized the CUDA Unified Memory (UM) mechanism and optimized the workflow for reading and training deep convolutional neural networks with ultra-high resolution images directly. The ultra-patch method has already gained prominent results [1] on both nasopharyngeal carcinoma (NPC) classification and colorectal cancer lymphoma metastasis classification. However, the generalization ability of the ultra-patch method and its performance on TMAs still remain unknown. By using NLST pathology dataset, we can further validate the generalization of the ultra-patch method.

[1] Training Deep Neural Networks Directly on Hundred-million-pixel Histopathology Images on a Large-scale GPU Cluster, https://sc19.supercomputing.org/proceedings/tech_poster/poster_files/rpost144s2-file3.pdf