Abstract
Heterogeneity in the tumor environment can be driven by multiple biochemical processes, such as immune infiltration, cell-to-cell genetic variation, and oxygenation gradients. Low oxygen, termed hypoxia, has been strongly correlated with metastasis, radiation resistance, and poor prognosis. One ongoing challenge is determining the magnitude of how hypoxia influences treatment response relative to other features, such as proliferation and micro vessel density. While it is possible to stain banked tissue for these markers, it is challenging to stain for endogenous markers of hypoxia. Exogenous probes have been developed to overcome this barrier; however, most clinical samples lack such probes. Previously, we have shown that spatial distributions of hypoxia correlate with morphological features of the tumor, such as necrosis and vasculature. Some of these features can be identified in hematoxylin and eosin (H&E)-stained sections, present for nearly all tumor samples. We hypothesize that a trained machine learning algorithm can be used to accurately predict hypoxia distributions given an H&E-stained section. Colorectal, pancreatic, and ovarian cancer cell line xenografts were cut and stained for markers of hypoxia, proliferation, perfusion, vasculature, and H&E. Binary masks of necrosis were generated, and together with the H&E and hypoxia images, were used to train a convolutional neural network. Once the model had been trained, a fairly precise prediction (MSE <0.012) had been obtained relative to the ground-truth hypoxia stained images. We believe a model trained on clinical H&E and hypoxia sections can be used to accurately predict hypoxia distributions from a H&E section. Predicted hypoxia distributions from existing H&E histology can be used to evaluate the influence of hypoxia on treatment outcome, by retroactively comparing a patient’s prognosis with levels of hypoxia. Correlations between hypoxia and prognostic measures such as tumor volume after irradiation, would indicate hypoxia may have modulated the efficacy of the treatment. Thus, computational stain prediction of hypoxia can serve as an invaluable tool in closing the gap between preclinical and clinical implementations of hypoxia-targeted therapies.
Citation Format: Mark Zaidi, Haotian Cui, Bo Wang, Trevor D. McKee, Bradly G. Wouters. Computational staining of tumor hypoxia from H&E images using convolutional neural networks [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-018.