Identifying the probability of genetic mutations in lung cancer using predictive and prognostic biomarkers from histopathological images

Abstract

Background: Lung cancer is the highest deadliest disease and second largest disease being diagnosed worldwide. In the age of precision medicine, determining a patient’s genetic status is critical. Finding the percentage of gene mutation of a particular biomarker will help in targeted therapy of a patient at an early stage. Objective: Histopathology images are larger in size which needs to be converted into smaller tiles for the computational purpose. Deep Learning Techniques could be applied on this huge number of histopathological images to derive the probability of gene mutation occurrence in predictive and prognostic biomarkers of lung cancer. Methods: In this work, a deep learning convolutional neural network (CNN) model (InceptionV3) is trained on histopathology images obtained from The Cancer Genome Atlas (TCGA) to accurately predict the mutated genes in lung adenocarcinoma. The convolutional neural network-based model predicts 10 major genetic mutations percentage, i.e., EGFR, FAT1, FAT4, KEAP1, KRAS, LRP1B, NF1, SETBP1, STK11, TP53. Results: InceptionV3 predicted the probability of gene mutation from the histopathology images and categorized the genes as predictive and prognostic. InceptionV3 yielded an accuracy of 82.36% and cross entropy of 37.62%. Conclusion: InceptionV3 was trained on histopathology images to predict gene mutations with an accuracy of 82%. Prediction of gene mutations with different CNN models like AlexNet and ResNet can be explored further.