RSNA2023 Leading Through Change
Daily Bulletin

Deep Learning Model for T2-Weighted Breast MR Can Slash Acquisition Times While Minimizing Noise

Monday, Nov. 27, 2023

By Evonne Acevedo

A novel deep learning (DL) sequence for T2-weighted breast MR images can reduce acquisition time by 59%—and employs super resolution that, in most cases, yielded sharper diagnostic images.



Caroline Wilpert, MD, a radiologist at the Clinic for Diagnostic and Interventional Radiology at University Medical Center Freiburg in Germany, presented findings from a prospective study involving 140 women who underwent 3T breast MR imaging with both a standard T2-weighted protocol and a DL-reconstructed sequence with a doubled acceleration factor, reduced phase resolution and identical spatial resolution.

With the DL sequence, acquisition time was reduced by 59%. Compared with a standard T2-weighted Dixon sequence, numerous image quality characteristics—including signal-to-noise ratio in breast structures and contrast-to-noise ratio in breast cancers—were improved with DL. "We did not find any additional artifacts in our sequences, and the subjective noise was lower in the DL sequence," Dr. Wilpert said.

The DL technique employs both a "de-noising" network and a super-resolution network, Dr. Wilpert explained. These networks were trained by an outside vendor on 25,000 images acquired at 1.5T and 14,000 acquired at 3T, from various body regions. An undersampled k-space dataset was used for image reconstruction.

"To enable gain in time, only half of the matrix was acquired for the DL network—305 by 576 versus 576 by 576—and acceleration was doubled for the DL sequence, with a factor of 4 instead of the standard 2," Dr. Wilpert said. "The de-noising network successfully compensated for the lower signal-to-noise ratio."

In addition, the investigators employed a super-resolution network, which interpolates image sequencing to produce a spatial resolution of 0.3 x 0.3 by 3.0 mm3 versus the standard 0.6 x 0.6 by 3.0 mm3, Dr. Wilpert explained.

"As a result, the images appear sharper," she said.

The team found it surprising that the image quality was so tangibly improved. "The effect of the reduced motion artifacts was most impressive, as it increased the value of the sequences," Dr. Wilpert said.

Also surprisingly, however, the opposite was true when it came to bone marrow. "That was an interesting finding we did not expect," Dr. Wilpert said. She suggested that bone marrow might be more susceptible to loss of structural details in the DW sequences because of its more heterogenous structure. This could be an effect of the edge enhancement combined with denoising, she said, emphasizing that the technique should not be used to evaluate bone marrow pathologies.

For breast imaging, Dr. Wilpert's team believes DL-based reconstruction could have a big impact on the efficiency of MR protocols.

"In times of limited scanner availability, both time and cost efficiency are becoming more important and should be addressed to meet the increased demand for breast MR imaging," Dr. Wilpert said. "In the future, a paradigm shift toward accepting lower image quality for the sake of ultrafast acquisitions could be conceivable. This could be partially compensated by the DL networks."

Access the presentation, "Reduced Acquisition Time of Deep-Learning Accelerated T2-Weighted Breast MR Imaging at 3T with Super Resolution: A Prospective Study on Image Quality," (S2-SSBR01-06) on demand at