How Neural Networks are used in Medical Imaging

When:
July 23, 2019 @ 6:30 pm – 9:00 pm America/Los Angeles Timezone
2019-07-23T18:30:00-07:00
2019-07-23T21:00:00-07:00
Where:
City: Santa Clara
Contact:
MAILTO:avid.farhoodfar@gmail.com

Co-sponsored by: dev_bhattacharya@ieee.org

Topic 1: Image Quality Assessment and Enhancement of Fetal Brain Magnetic Resonance Imaging using Convolutional Neural Networks

(Sayeri Lala)

Fetal Brain Magnetic Resonance Imaging (MRI) complements Ultrasound in diagnosing brain pathologies. Despite the utility of fetal brain MRI, the substantial, rapid, and non-periodic nature of fetal motion can severely degrade image quality, often requiring several repeat acquisitions in the clinic.

To improve scan efficiency and diagnostic quality, we investigated Convolutional Neural Networks (CNNs) to perform rapid image quality assessment and enhancement. In a retrospective study, we demonstrated that a CNN for image quality assessment achieves good performance, and observed from saliency maps that the classifier learns relevant criteria. We also showed that on retrospectively undersampled data, a CNN based architecture produces higher quality images compared to traditional reconstruction algorithms. The fast evaluation and reconstruction times of the CNNs suggests their potential to be used during the scan to monitor and enhance image quality.

Topic 2: Deep learning and GAN enable 4x faster scans for clinical radiology (MR, PET and more) (Enhao Gong)

Abstract: 

In this talk, Dr. Gong will introduce the research and technology development at Stanford and Subtle Medical on AI-powered medical imaging reconstruction and enhancement. Subtle Medical develops AI/Deep Learning software solution that recently gains FDA clearance and enables faster & safer radiology exams. For faster PET and MRI exams, deep learning algorithms are used to replace the conventional iterative optimization based algorithm. Using ResNet-based recurrent structure and generative adversarial network (GAN) based adversarial loss, the model successfully reconstructs images from low-quality into high-quality with significantly better signal-to-noise ratio and resolution information. Results from clinical partners and deployment sites such as Hoag Hospital and UCSF demonstrate the improvement of clinical workflow and diagnostic quality. The AI inference in deployment is further accelerated with industry frameworks such as NVIDIA tensorRT and Intel OpenVINO. The solutions are currently in clinical usage and commercialization, providing immediate and quantifiable values of AI to hospitals and imaging centers.

In further research work, task-driven GAN and cycle-GAN are used to further improve the performance of DL-based medical image enhancement, demonstrating superior performance for more extreme application of ultra-low-dose PET and 1min-MR scan that will disrupt how radiology is used in healthcare.

Speaker(s): Ms. Sayeri Lala, Mr. Enhao Gong

Agenda:

Sayeri Lala  will present for 20 min and Enhao Gong present for 40 minutes.

Location:
Bldg: Nvidia Buliding E
2800 Scott Blvd
Santa Clara, California
95050