Application of automated segmentation and classification of PD-L1 immunohistochemistry images in cancer treatment prediction.

Abstract

e15572 Background: Therapies targeting immune checkpoints, such as the programmed cell death-1 (PD-1) / programmed cell death ligand 1 (PD-L1) pathway, have become successful in treating several cancer types, particularly those with high PD-L1 expression. Patient response rates, however, are variable, demanding a more sensitive companion diagnostic test than the current method of manual scoring of PD-L1 immunohistochemistry images. The low predictive value of manual PD-L1 scoring is at least partially compounded by the pathologist-to-pathologist or hospital-to-hospital variation. To eliminate the variation caused by human factors, we developed an alternative PD-L1 scoring method based on automated image analysis and assessed its predictive and prognostic values. Methods: Archival or fresh tumor biopsies were analyzed for PD-L1 expression by immunohistochemistry. Digital images were automatically scored as PD-L1 positive or negative using a custom algorithm implemented in Python and Tensorflow, which uses a fine-tuned, ImageNet pre-trained, Inception-ResNet-V2 model as the binary classifier along with a modified U-net image segmentation model to determine the final prediction. Supervised machine learning was used to control for heterogeneity introduced by PD-L1 positive inflammatory cells in the tumor microenvironment. Samples from 256 patients were collected and used in algorithm training and verification by random assignment with an 8:2 split ratio. Results: In total, 10,000 images with balanced PD-L1 positive and negative distributions were obtained. A highly significant correlation of PD-L1 scores was found between pathologist-based consensus reading and automated analysis using our custom algorithm (R = 0.97, p < 0.0001). The automated analysis reached 0.94 sensitivity and 0.96 specificity. Additionally, it showed excellent reproducibility (R = 1.0, p < 0.0001) in independent trials executed on different workstations, which is far superior to the manual scoring performed by human examiners. Conclusions: Our automated PD-L1 evaluation algorithm significantly reduces scoring variability. It may help to identify cancer patients who can have optimal response to PD-1/PD-L1 related immune therapy and therefore facilitate patient stratification in clinical practice.