Neural network assesses liver tumor burden on MRIs

Last Updated: 2020-08-14

By Reuters Staff

NEW YORK (Reuters Health) - For management of neuroendocrine tumors, a deep-learning algorithm can provide MRI interpretations comparable to those of humans, researchers report.

Three-dimensional volumetric tumor assessment has superior accuracy for assessing change in tumor size and burden, Dr. Alexander Goehler of Harvard Medical School, in Boston, and colleagues note in the Journal of the American College of Radiology.

Although there have been some proposals seeking to automate the process, they add, this time-intensive procedure is used only in small clinical trials and is not part of routine clinical care.

To investigate the feasibility of an artificial-intelligence approach, the researchers studied data on 64 patients with neuroendocrine tumors who underwent at least two consecutive liver MRIs with gadoxetic acid. That agent was chosen because of its favorable contrast properties. None of the patients had had interim ablative, surgical therapy, or other focal therapy such as stereotactic radiation therapy.

Most patients had their primary tumor in the small bowel and the median time between the two examinations was four months.

The researchers then developed a 3-D neural network using a U-Net architecture with ResNet-18 building blocks. This first detected the liver and then lesions within the liver. The deep-learning algorithm developed detected liver metastases, and then assessed the interval change in tumor burden between two multiparametric liver MRI examinations.

Two experienced board-certified radiologists manually segmented the liver and lesions using ITK-SNAP on all relevant images within the 3-D data sets, and subsequently reviewed each other's work. Discrepancies were resolved by consensus.

The team was able to demonstrate a high concordance (91%) between the deep-learning algorithm and manual assessments by the clinical radiologists with regard to interval change in neuroendocrine hepatic metastatic tumor burden.

In classifying liver segments as being diseased or healthy, the algorithm had a sensitivity of 0.85 and a specificity of 0.92. The sensitivity per lesion was 0.84.

One of the study's limitations is that the data come from a single institution and there were only 128 liver examinations, the authors note.

"Our algorithm displayed high agreement with human readers for detecting change in liver lesions on MRI, offering evidence that artificial intelligence-based detectors may perform these tasks as part of routine clinical care in the future," they conclude.

Dr. Goehler did not respond to requests for comments.

SOURCE: https://bit.ly/31SGATN Journal of the American College of Radiology, online July 25, 2020