Rick Wray

M.D.

UC Davis Health

Assistant Professor of Radiology

PLCO (Learn more about this study)

PLCO-378

Jul 5, 2018

Effective Detection of Chest Pathophysiology Utilizing Modern Methods in Machine Learning

Currently in the field of modern radiology, radiologist must manually review chest X-rays (CXR) and computed tomography (CT) scans in order to diagnose chest and lung pathophysiology. However, recent advances in machine learning (ML) promise to give doctors tools that can assist in the diagnosis of disease. As our understanding of these diseases increases and their various treatment modalities increase, so does the need for greater specificity in diagnostic accuracy. Recently released datasets of CXR pathophysiology, such as ChestX-ray14, have been particularly useful in improving state of the art CXR diagnostic assistants and decision support systems. We intend to improve existing CXR diagnostic models by utilizing a large dataset of CXRs from the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening Trial, and by utilizing advanced ML methods. We propose a focused, and minimal risk research project to achieve this goal. Initially, our project will utilize the PLCO CXR dataset to serve as a "pre-training" dataset that will be leveraged to teach our ML model to recognize certain features specific to CXR imagery. After which, we intend to "fine-tune" and improve our pre-trained model on the Chest-Xray14 dataset. It is our hope that with the extensive pre-training offered by the PLCO dataset, combined with our utilization of modern and novel methods for performing ML image recognition, that we will be able to achieve state of the art results for CXR pathophysiology recognition. Once we determine the best modeling approach for identifying CXR pathophysiology with Chest-Xray14, we will then create a new model that is pre-trained on Chest-Xray14, and fine-tuned and validated with PLCO CXRs. In this respect we hope to achieve state of the art pathophysiology detection results for both datasets. Our project is minimal risk because no patient identifying information will be used in the course of this study, nor will we make an intervention in treatment of care. Our proposal is to improve existing ML models, with new methods and existing datasets. We believe that our advances towards improving the performance of radiologic diagnostic assistants will assist the efforts of clinicians, and improve the efficacy of future healthcare systems both in the United States and globally.

AIM 1: Improve existing state of the art methods for detecting lung pathophysiology via CXR imagery
• 1A. Utilize PLCO CXR dataset to “pre-train” our machine learning models so that our models can learn the underlying topology and potential heterogeneity of a chest X-ray (CXR)
• 1B. Implementing previous methodology, "fine-tune” our machine learning models so that our models will learn to detect specific pathophysiology present in ChestX-ray14 dataset. Benchmark performance results for our models
• 1C. Augment machine learning models creating in sub-aim 1B with state of the art methods, such as model ensembling, and attention mechanisms for improving image recognition models.

AIM2: . Once the best modeling methodology for CXR pathophysiology recognition has been found, we will create an improved model that is pre-trained on Chest-Xray14 and then our final model will be validated on the PLCO CXR dataset.