Deep Learning Shows Predictive Power in COPD Patients

Monday, Nov. 26, 2018

A deep learning (DL) algorithm can discern features of chest CTs and potentially predict the prognosis for patients with chronic obstructive pulmonary disorder (COPD), according to research presented Sunday.



Imaging is frequently needed for patients with COPD, which affects more than 16 million Americans and 251 million people worldwide, as many develop a host of respiratory problems.

Researchers lead by Tara A. Retson, MD, PhD, a first-year resident at the University of California in San Diego, studied the potential of DL — a type of artificial intelligence — to glean information from a chest CT scan that might predict the course of the disease; predictions for outcomes like hospitalization, chronic bronchitis and mortality that could then be compared with a group of COPD patients who were followed for five years. Deep learning processes information through convolutional neural networks, which are designed like a series of layers that get progressively more complicated.

"With deep learning, we can give the algorithm a scan and it comes up with features that it thinks are important, like the percentage of emphasema," Dr. Retson said. "It's pretty revolutionary. We used to have to provide every single detail, and now the algorithm is able to figure it out on its own."

Dr. Retson tested the algorithm this past summer on CT scans for 160 COPD patients from the San Diego area who were enrolled in the COPDGene® multicenter study examining the underlying genetic factors of the disease. She trained the algorithm on 20 chest CT scans with assessment by five-fold cross validation, an approach that divides data into five pieces and runs it five times to determine average accuracy.

Dr. Retson is in the process of studying the algorithm on a much larger data set consisting of 10,000 images from the COPDGene program. She has evaluated the algorithm on 2,000 of those images, using 80 percent of the images to train the algorithm and the remaining 20 percent to test it.

So far, the results show a strong correlation between the algorithm's predictions and the actual outcomes.

"These results are pretty solid, especially in terms of predicting total lung capacity and percent of emphysema," Dr. Retson said. "Now we want to make some technical adjustments and enlarge the study."

Algorithm Tested on Larger Scale

The algorithm could present an improvement over existing methods since it may provide information on how the disease is progressing.

"Our ultimate goal is to develop predictive measures for COPD we can take from the algorithm and give to clinicians," Dr. Retson said.

While the algorithm needs further development before it can assign patients a specific percentage of risk for future outcomes like chronic bronchitis or hospitalization, Dr. Retson noted that the measures inferred from the scans, including total lung capacity, emphysema and functional residual capacity, have previously been correlated with patient outcomes.

"Because of this, something these measures can do is help physicians understand, track, and make informed decisions about a patient's health and likely trajectory," she said. "As an added benefit, clinically making the same measurements requires advanced testing or manual image analysis, so automation can ideally save time and money."

Convolutional neural networks (CNN) are capable of accurately inferring pulmonary measurements from chest CTs by automatically identifying the image features most important for determining outcomes. These two patients had similar manually calculated percentages of emphysema. The algorithm "underestimated" emphysema on the right image, possibly due to the large bullae, a structure it may not have classified as lung. Left: Predicted emphysema: 25.01 percent; calculated ground truth: 24.43 percent. Right: Predicted emphysema: 10.90 percent; calculated ground truth: 25.73 percent.