Leveraging artificial intelligence for early detection of lung cancer

Predictive models developed by an interdisciplinary U-M research team have improved early lung cancer detection beyond traditional measures, with the potential to save lives.

Lung cancer, like many cancers, is much more treatable and less fatal if caught early––the 5-year survival rate for patients with early stage lung cancer is 63.7%, compared to just 8.9% for those with advanced disease. New techniques developed by an interdisciplinary team of researchers in Electrical and Computer Engineering (ECE) and the U-M Medical School are improving current methods for the early detection of lung cancer.

The current standard for preventative lung cancer screening is serial low-dose computed tomography (LDCT) scans. These annual scans are recommended for adults 50–80 years old who are at risk of lung cancer due to a history of cigarette smoking. However, at the earliest stages, lung cancer is generally asymptomatic and any growths of abnormal tissue, called nodules, are too subtle to determine whether they are benign or malignant from a baseline LDCT scan. Radiologists compare the progression of any observed nodules, or growths of abnormal tissue, within the lungs to label them as benign or to order further diagnostic tests such as biopsies.

A patient in a hospital gown lies flat on the bed of a CT scanner, a large piece of medical equipment shaped like a donut.
Patient undergoing CT scan test. Photo: Wavebreak Media Premium / Alamy Stock Photo

“Given the nature of the early stage of lung cancer, we believe that as the disease progresses biologically over time, focusing on single CT scans can not accurately reflect the overall risk,” said ECE PhD student Yifan Wang. “To align with the natural history of the lung nodule and achieve early diagnosis, we aim to look forward, rather than just reflecting on the past or focusing solely on the present, to find the answers.”

To do this, Wang, his PhD advisors Lei Ying and Chuan Zhou, and their team of collaborators, have developed predictive models to aid in the early diagnosis of lung cancer prior to further screenings. Rather than making decisions from visual inspection of a nodule at a single time point or from risk factors such as age and sex, Wang’s models leverage serial LDCT data to predict the future progression of small nodules that appear benign.

The team created a radiomics-based reinforcement learning model called S-RRL to predict the risk of malignancy for nodules, trained and tested on existing LDCT data from 2500 total participants in a National Lung Screening Trial. They found that the S-RRL model correctly predicted nodule risk of malignancy better than the standard clinical diagnostic models, known as the Brock model and Lung-RADS.

Building on this work, Yifan also developed a deep predictive model called GP-WGAN to forecast nodule growth patterns as AI-generated images. The study showed that the images predicted by GP-WGAN aligned well with the real follow-up LDCT images, and that the use of the predicted images from the model during baseline screening improved diagnostic accuracy.

Two 3-by-3 grids of CT scan images, showing white shapes of lung nodules on a black background.
Example images of six lung nodules on the baseline LDCT scans, the 1-year follow-up LDCT scan, and the GP-WGAN predicted nodules. Adapted from Wang et al.; Cancers Volume 16, Issue 12, June 2024.

“The application of AI in healthcare is actually very unique, and we must adopt a distinct perspective from the conventional approach used for general AI applications in daily tasks,” explained Wang. “When models are deployed in clinical settings, they demand extensive validation due to the minimal tolerance for errors. To the best of my knowledge, our study is the first to employ this approach for solving the problem of early diagnosis of lung cancer.”

Although there is much more work to be done before these models can be used by clinicians in a healthcare setting, they have the potential to streamline the diagnostic process at baseline lung cancer screenings, enabling doctors to tailor care plans to their patients, avoid unnecessary invasive procedures, and improve their prognosis.

Future work by members of the team includes exploring additional algorithms for early lung cancer diagnosis, such as the EarlyStop-RL algorithm and the use of diffusion models, refining the current GP-WGAN model, and incorporating multimodal risk factors like patient demographics into the existing models. Lei Ying is a professor of ECE and Chuan Zhou is a research professor of Radiology. Other collaborators on the project are Profs. Elizabeth Lee and Lubomir Hadjiyski of Radiology; Prof. Emeritus Heang-Ping Chan of Radiology; Prof. Ella A. Kazerooni of Radiology and Internal Medicine; and Aamer Chughtai, Diagnostic Radiologist at the Cleveland Clinic.

This study is supported by NIH grant U01CA216459.

Explore:
Cancer; Lei Ying; Machine Learning; Research News; Signal & Image Processing and Machine Learning; e-HAIL