Deep Learning: Automated PD-L1 Tumor Cell Scoring of Resected NSCLC
PD-L1 expression measured by immunohistochemistry helps identify non-small cell lung cancer (NSCLC) patients that may respond to anti-PD-1/PD-L1 immunotherapies. In this webinar, our featured speaker will present a novel deep learning solution for the automated scoring of PD-L1+ tumor cells (TC) in whole slide images of resected NSCLC.
A convolutional neural network (CNN) was used for the fine-grained classification of tissue regions into three classes: (1) regions of membrane-positive epithelial tumor cells [TC(+)], (2) regions of membrane-negative epithelial tumor cells [TC(-)], and (3) other regions that could wrongly influence scoring, i.e. macrophages, positive and negative lymphocytes, stroma and/or necrosis. The TC score was calculated as the ratio of the area of the classified TC(+) region to the sum of the areas of the classified TC(-) and TC(+) regions.
Two sets of ~225k training and ~30k testing patches (128×128 pixels) were created from manual partial annotations from (N=22) train and (N=12) test slides (Ventana-SP263). Training a fully convolutional network yielded maximum accuracy of 0.89 on test patches. The trained network was applied on (N=433) unseen confirmation slides and the TC score calculated for each slide based on the classified TC(+) and TC(-) regions. A non-linear gamma mapping to the manual TC scores by a trained pathologist was then estimated to maximize Overall Percent Agreement (OPA) at ≥25% criterion using two-fold cross-validation.
Evaluation against pathologist scoring on the confirmation slides yielded higher Overall (OPA), Negative (NPA) and Positive (PPA) Percent Agreement values at ≥25% criterion, higher Lin’s correlation and lower mean absolute error than a baseline approach relying on a heuristic detection of individual epithelium cells [ESMO-2017-103P]. Scoring by a second pathologist was available on a subset (N=170) of the confirmation slides. Average and standard deviation results on this subset confirm the above observations and suggest that the approach is getting close to inter-pathologist variability.
Join this webinar to learn how using deep learning to identify PD-L1 positive and negative tumor cell regions enables the automated scoring of PD-L1 TCs at the ≥25% expression level in resected NSCLC. These findings should be confirmed with additional tumor sets.