Deep Learning and Automation: Research from Ci2 Visiting Scholars

TweetShareEmail
By UCSF Ci2 Team

Visiting scholars offer institutions an opportunity to diversify educational perspectives and learn new approaches. The University of California, San Francisco (UCSF) Center for Intelligent Imaging (ci2) was pleased to welcome two scholars this summer, Yassine Guennoun and Rodrigue Reibel, from École Polytechnique University in Palaiseau, France. Guennoun and Reibel presented their research at the August SRG Pillars Meeting.

Yassine Guennoun, working in the Rauschecker-Sugrue Lab at UCSF, presented "Meningioma segmentation and uncertainty quantification," at the August 18 meeting. His research focuses on applying deep learning to medical imaging, with an emphasis on uncertainty-aware methods for brain tumor segmentation and volumetric analysis. Guennoun collaborated with several researchers, including Pierre Nedelec, MS, Junchi Wei, MS, Evan Bloch, BS, Mark McArthur, MD, and Andreas Rauschecker, MD, PhD.

Current models in meningioma segmentation provide accurate tumor segmentation but can't express how confident they are. Guennoun proposed using logits and softmax, which prevent meaningful uncertainty estimation and hide ambiguous regions. The dataset included 1,324 trainings scans and 331 validation scans. Guennoun applied random spatial and intensity augmentations to these scans, testing against 93 scans from 55 patients. 

Deployment objectives for this research include enabling clinical applications — 90 seconds per scan for predictions, followed by a post-processing pipeline to generate volume plots for neuro-radiologists at UCSF. Clinical deployment would mean providing radiologists with a confidence interval, allowing for more confident assessment of tumor growth or stability.  

Rodrigue Reibel, working in the Rauschecker-Sugrue Lab at UCSF, presented "Automated Identification of White Matter Tract intersections for Deep Brain Stimulation (DBS) targeting." His work focuses on developing and applying machine learning methods to neuroimaging data to investigate how brain structure and connectivity relate to functional activity and cognition. Reibel was advised by Leo Sugrue, MD, PhD, and supervised by Pierre Nedelec, MS, on DBS for treatment-resistant depression.

Normative and subject-specific targeting are some time-consuming and non-reproducible methods for DBS targeting. To improve upon these methods, Reibel designed an automated pipeline for subject-specific identification of tract intersection. This method of targeting is scalable, reproducible, not operator-dependent and relies on single-subject tractography. 

Reibel sampled a dataset of 12,000 children, tracking brain, behavior, genetics and environment. The scans from this dataset found a strong intersection when scored on the heat map variability, showing evidence of white matter variability. Deployment opportunities included using fMRI to assess connectivity at the target location and direct targeting leveraging deep learning models like TractSeg to predict intersection scores on the heat map. 

To learn more about the upcoming SRG Pillar meetings, visit the ci2 events page.