arXiv preprint 2022

SoftDropConnect

SoftDropConnect (SDC) – Effective and Efficient Quantification of the Network Uncertainty in Deep MR Image Analysis

Qing Lyu, Christopher T. Whitlow, Ge Wang

Current or former lab member. Select a name to see all of their projects.

Method overview of SoftDropConnect
Dropout (left), DropConnect (middle) and SoftDropConnect (right). In Dropout and DropConnect the drop mask weights are either one or zero; in SoftDropConnect they are continuous (Fig. 1 of the paper).

Abstract

SoftDropConnect (SDC) is a simple Bayesian inference method for quantifying the uncertainty of a neural network in medical imaging. Dropout and DropConnect disable parameters at random. SDC instead modulates them continuously during training and testing, so the affected channels keep operating.

SDC was tested on glioma segmentation and brain metastasis classification. Compared with Bayes by Backprop, Dropout and DropConnect, it improved whole-tumor segmentation Dice by 3.4%, 2.5% and 6.7%, and metastasis classification accuracy by 11.7%, 3.9% and 8.7%, while greatly reducing epistemic and aleatoric uncertainty.

What’s in the repository

  • Code for brain tumor segmentation
  • Code for a digit recognition experiment

Citation

@article{lyu2022softdropconnect,
  title   = {SoftDropConnect (SDC) – Effective and Efficient Quantification of the Network Uncertainty in Deep MR Image Analysis},
  author  = {Lyu, Qing and Whitlow, Christopher T. and Wang, Ge},
  journal = {arXiv preprint arXiv:2201.08418},
  year    = {2022}
}
Back to projects