Imaging Based Models to Improve Lung Cancer Diagnosis

Abstract

Per the American Cancer Society, lung cancer is the second most common cancer in both men and women, and the leading cause of cancer death, making up almost 25% of all cancer deaths. As such, it is pivotal to better detect and locate lung cancer from chest radiographs (x-rays) and computed tomography (CT) scans, as well as accurately estimate the risk of future lung cancer, all while factoring in the importance of model explainability in clinical settings.

Recent advances in deep learning have led to increased applications of machine learning to medical imaging. In this work, we seek to better understand lung cancer through the computer vision tasks of risk prediction, localization, and incorporating image priors across the clinical imaging modalities of chest radiographs and computed tomography scans. The task of lung cancer tumor risk prediction allows the model to identify current cancers and act as an objective second reader, or help radiologists flag risky exams. The task of localization of both present and future cancers can help radiologists ascertain current and future regions of interest that should be further examined. The task of incorporating image priors allows the risk prediction model to mimic the radiologist screening workflow of using multiple screening images when available.

We develop our methods on chest radiograph data from the National Institute of Health (NIH) dataset and low dose computed tomography (LDCT) data from the National Lung Screening Trial (NLST) dataset. We study the aforementioned three tasks across these two imaging modalities and explain how these tasks can improve patient care in clinical settings. Our results show that models based on LDCT can accurately detect current cancers but also provide longer term risk assessment beyond what can be achieved using risk factors alone. Through this work we aim to improve clinical care, offering new tools to improve patient outcomes.