Scientists from the UCSF Ci2 Propose a Deep Learning Framework to Address the False Negative Problem of MRI Reconstruction Networks

High data quality is a priority in medical image analysis. Magnetic resonance imaging (MRI) has the capability of satisfying such requirement when it comes to screening soft tissues. Nevertheless, MRI has a limitation of requiring long scanning time. Deep learning models have been shown to be successful in accelerating MRI reconstruction, over traditional methods. However, it has been observed that these methods tend to miss rare small features, such as meniscal tears, and subchondral osteophytes, in musculoskeletal applications. Such findings are concerning for radiologists, as these small and rare features are the particularly relevant in clinical diagnostic settings.

To address these concerns, a group of scientists from the UCSF Center for Intelligent Imaging (ci2) proposed a framework to find the worst-case false negatives by adversarially attacking the trained models and improve the models’ ability to reconstruct the small features by robust training. Kaiyang Cheng from the Department of Electrical Engineering and Computer Sciences, University of California, Berkeley was first author. Authors from UCSF included Francesco Caliva, PhD, Rutwik Shah, MD (postdoctoral scholars); Misung Han, PhD (assistant researcher) and Valentina Pedoia, PhD and Sharmila Majumdar, PhD, both faculty members in the UC San Francisco Department of Radiology and Biomedical Imaging.

This work won the 2020 Best Paper Award at Medical Imaging with Deep Learning (MIDL) earlier this month. The award recognizes the highest quality full-length paper presented at the conference. The focus lies on novel methodological concepts with great potential of medical impact. You can access the paper, reviews and video presentation online or below.  

The UCSF Ci2 was founded in fall 2019 as a hub for the multidisciplinary development of AI in imaging to meet unmet clinical needs and provide a platform to measure impact and outcomes of this technology.

Related Content

Related People