Workshop: Machine Learning Algorithms in Histology and Radiology for Cancer Drug Discovery and Development
Abstract: Background: Lung cancer is one of the most common cancers in the world. It is a leading cause of cancer death in men and women in the United States. Computational approaches such as deep learning could help accurate and efficient analysis of biomarkers, both histopathology and radiology, to assist cancer drug discovery and development. PD-L1 expression in the tumor micro-environment is one of the efficacy predictors of cancer treatment using immune checkpoint inhibitors. Tumor proportion score (the ratio of PD-L1 expressed area within the total tumor area) is often used as a measure to distinguish low and high PD-L1 expression. However, due to difficulty in quantitative measurement of microscopic images, the proportion score is prone to be evaluated subjectively. A more objective way by using computational techniques which can quantify PL-L1 expression only in the tumor cells is highly desirable. In addition, nodules detected on a chest radiograph (CR) is a good clinical marker for detection and progression of the disease. Interpretation of CR has problems of overlooking nodules and the reported error rates for missed lung cancer on CR are 20%-50%. Deep learning techniques can overcome the accuracy challenges of conventional interpretation and can improve the quality of the clinical biomarker.
Objective: One of our objectives is to create an image analysis system to automatically recognize the “total tumor area” and “tumor area with PD-L1 expression” in order to derive a quantitatively measured “tumor proportion score”. The second objective is to develop an algorithm with deep learning techniques for detecting lung cancer on CR and assess its performance.
Materials and Methods: For the PD-L1 expression study, whole slide images (WSI) of lung cancer pathology were used. Small patches were randomly selected from WSI to train our segmentation model. Deep convolutional neural networks were built and the segmentation model was trained using the extracted and annotated patches. Training has been done with 34 patches of 1000 x 1000 pixel images. The algorithm for lung nodule detection was developed by segmentation method with encoder-decoder architecture. The encoder used Inception-ResNet-v2 model as the backbone algorithm to extract features. The decoder was designed to compute up convolution to detect lung nodule or mass with segmentation. Training and test data were obtained from University of Osaka Medical School.
Results and Discussions: We have demonstrated that application of deep learning to microscopic images has potential in the following tasks: Distinguishing cancerous regions from other tissue regions, Segmentation of regions with PD-L1 expression, Quantitative evaluation of tumor proportion score. An accurate and fast quantitative evaluation of PD-LI expression on tumor cells (not confounded by PL-L1 expression in non-cancer area) would provide a more accurate indication of drug efficacy in high content screening. The algorithm for detection of nodules in CR achieved high performance. The sensitivity, Positive Predictive Value (PPV), and False Positive per Image (FPI) were determined to be 73.0%, 85.3%, and 0.13, respectively. This lung nodule detection algorithm could provide a reliable radiology biomarker in lung cancer clinical trials.