Precision radiology may become possible with deep learning-based abdominal CT segmentation
A deep learning algorithm developed by researchers at the Mayo Clinic in Rochester, Minnesota, segmented abdominal CT images to determine body composition similarly to, and at times, better than trained radiologists.
In the study, published online Dec. 11 in Radiology, researchers created an algorithm based on U-Net architecture and trained it to perform abdominal segmentation on a dataset of 2,430 2D CT exams. The algorithm was tested on 270 CTs and a separate set of 2,369 patients with hepatocellular carcinoma (HCC). All exams were performed between 1997 and 2015 on patients with an average age of 67 years old.
To establish a ground truth, an expert radiologist used a semiautomated approach to segment into four areas: subcutaneous adipose tissue, muscle viscera and bone along with pixels external to the body.
When compared to the reference segmentation, the algorithm “met or exceeded” the expert’s manual effort, according to first author Alexander D. Weston with Mayo Clinic’s Department of Biomedical Engineering and Physiology, and colleagues. The model achieved Dice scores (mean ± standard deviation) of 0.98 ± 0.03, 0.96 ± 0.02, and 0.97 ± 0.01 in the test set, and 0.94 ± 0.05, 0.92 ± 0.04, and 0.98 ± 0.02 in the HCC data set, for the subcutaneous, muscle, and visceral adipose tissue compartments, respectively.
“Our results suggest that an accurate 3D segmentation is possible by using a simple 2D model, which could vastly reduce both the complexity and the amount of training data required to develop a segmentation tool,” Weston and colleagues wrote.
Body composition can play an important role in clinical diagnosis, but calculating such measurements is time consuming. Certain composition markers have also been shown to predict surgical outcomes. The method developed in this study could better diagnose other metabolic conditions, Weston et al. wrote, but would require a study involving larger populations.
Paul Chang, MD, of the University of Chicago held out similar hope for the deep learning method, noting such machine learning applications could help the field enter into an era of “precision radiology" in a related editorial.
Chang described body composition as “typical of precision radiology” in that it adds value to many cardiovascular, oncologic and surgical outcome conditions. And despite the study’s challenges, including relatively small training and test cases, Chang held the findings of Weston et al. in high regard.
“Challenges notwithstanding, this study is a good example of how machine learning can and should be applied to radiology,” Chang wrote. “It is not to replace radiologists but to enable us to provide precision radiology in an efficient and scalable manner for our patients.”
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.