Lung cancer is the leading cause of cancer death. In order to increase survival, therapies are urgently needed to intercept the cancer development process and decrease the rate of patients presenting with advanced disease. A potential promising point of interception is to develop therapies to reverse or delay the development of lung premalignant lesions (PMLs). About 20% of lung cancers arise in the epithelial layer of the bronchial airways and these are preceded by the development of PMLs that are important clinical indicators of lung cancer risk in the airways or at remote parenchymal sites. As part of the NCI-Moonshot our group is engaged in creating a multi-omic lung Pre-Cancer Atlas (PCA). The success of this project in creating clinically relevant biomarkers and therapeutics depends on accurate assessments of histology and immune infiltrates in PMLs. Currently, however, pathologic assessment of the morphological stages of increasing abnormality from hyperplasia, metaplasia, dysplasia (mild, moderate, and severe), to invasive carcinoma is challenging and not routine. The objective of the proposed study is to develop and disseminate a computationally efficient deep learning framework to annotate a variety of histologic features in PMLs from the Lung PCA and associate these features with clinical and genomic data. Our central hypothesis is that deep learning can be applied to digitized H&E whole slide images (WSIs) of bronchial PMLs to identify a comprehensive set of histologic features and metrics summarizing their spatial organization that may enhance biomarkers of PML progression to cancer. We will test this hypothesis by pursuing two specific aims. First, we will annotate PMLs and develop a semantic segmentation framework using deep learning to predict histologic features of PMLs. Second, we will disseminate our deep learning framework and show its utility in enhancing PML-associated biomarkers. The proposed study is significant because the framework we develop can be applied to predict other features in the WSIs from PMLs and be modified to encompass other PMLs of the lung (e.g. those associated with lung adenocarcinoma) as well as other premalignant lesions found in other epithelial tissue types such as breast, colon, prostate, etc. Currently, deep learning approaches have not been applied to PMLs, and this proposal is innovative in the unique clinical specimens that it leverages with corresponding genomic and clinical data and in its development of a patch- based convolutional neural network to predict histologic features of PMLs. Our long-term goal is to develop a deep learning framework to predict a variety of features from lung PML WSIs and integrate these with genomic data on these same samples to discover robust biomarkers of PML progression and therapeutics to prevent invasive cancer development.
The proposed study aims to develop and disseminate a computationally efficient deep learning framework to annotate histologic features in lung squamous premalignant lesions using digitized H&E whole slide images and associate these features with clinical and genomic data. We also aim to demonstrate the utility of our method in providing image-based features to enhance our genomic biomarkers of premalignant lesion progression. The methods and results based on the whole slide images can be integrated into a larger lung premalignant lesion atlas building initiative to advance lung cancer interception.
