SUD: Supervision By Denoising for Medical Segmentation

Speaker: Sean I. Young

Date: Tuesday, November 02, 2021

Time: 4:00 PM to 5:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: 32-507

Event Type: Seminar

Room Description:

Host: Polina Golland, CSAIL

Contact: Polina Golland,

Relevant URL:

Speaker URL: None

Speaker Photo:

Reminders to:

Reminder Subject: TALK: SUD: Supervision By Denoising for Medical Segmentation

Training a fully convolutional network (FCN) for semantic segmentation
typically requires a large labeled dataset with little label noise if
good generalization is to be guaranteed. For many segmentation tasks,
however, data with pixel- or voxel-level labeling accuracy are scarce
due to the cost of manual labeling and variability in the labeling
even across domain experts. Therefore, training segmentation networks
to generalize better by additionally leveraging the abundance of
unlabeled data is a problem of both practical and theoretical
interest. In this work, we propose to train a segmentation network
further on unlabeled data using the denoised version of the network
output itself as a soft segmentation target. Our proposed
semi-supervised training procedure facilitates incorporation of a
range of non-differentiable denoising filters while guaranteeing
locally convergent training. To take full advantage of our so-called
“supervision by denoising” framework, we propose in tandem a number of
filtering techniques to denoise corrupted intermediate segmentation
outputs, which outperform conventional mathematically crafted
denoisers and simpler regularization approaches. We apply “supervision
by denoising” to the training of two segmentation
FCNs—three-dimensional brain segmentation and cortical surface
parcellation—to demonstrate the utility of the proposed framework.

Research Areas:
Graphics & Vision

Impact Areas:
Health Care

See other events that are part of the Biomedical Imaging and Analysis 2021 - 2022.

Created by Polina Golland Email at Wednesday, October 13, 2021 at 9:09 PM.