Abstract
Pathology images capture tumor histomorphological details in high resolution. However, manual detection and characterization of tumor regions in pathology images is labor intensive and subjective. Using a deep convolutional neural network (CNN), we developed an automated tumor region recognition system for lung cancer pathology images. From the identified tumor regions, we extracted 22 well-defined shape and boundary features and found that 15 of them were significantly associated with patient survival outcome in lung adenocarcinoma patients from the National Lung Screening Trial. A tumor region shape-based prognostic model was developed and validated in an independent patient cohort (n = 389). The predicted high-risk group had significantly worse survival than the low-risk group (p value = 0.0029). Predicted risk group serves as an independent prognostic factor (high-risk vs. low-risk, hazard ratio = 2.25, 95% CI 1.34–3.77, p value = 0.0022) after adjusting for age, gender, smoking status, and stage. This study provides new insights into the relationship between tumor shape and patient prognosis.
Highlights
Lung cancer is the leading cause of death from cancer, with about half of all cases comprised of lung adenocarcinoma (ADC), which is remarkably heterogeneous in morphological features[1,2] and highly variable in prognosis
Utilizing the tumor shape features extracted from the pathology images in the National Lung Screening Trial (NLST) dataset, we developed a prognostic model to predict patient survival outcome
The pathology image was divided into 300 × 300 pixel image patches, which were classified into tumor, non-malignant, or white categories using a convolutional neural network (CNN) model
Summary
Lung cancer is the leading cause of death from cancer, with about half of all cases comprised of lung adenocarcinoma (ADC), which is remarkably heterogeneous in morphological features[1,2] and highly variable in prognosis. Studies have shown that morphological features are associated with patient prognosis in lung cancer as well[4,5,7]. Deep learning methods, such as convolutional neural. For analysis of H&E-stained pathology images, deep learning methods have been developed to distinguish tumor regions[14], detect metastasis[15], predict mutation status[16], and classify tumors[17] in breast cancer as well as in other cancers. Due to the complexity of lung cancer tissue structures (such as microscopic alveoli and micro-vessel), deep learning methods for automatic lung cancer region detection from H&E-stained pathology images are not currently available. Automatic tumor region detection in pathology images allows us to better characterize tumor region boundaries and extract tumor shape and boundary-based features
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
