0%

(MICCAI 2018) DeepASL:Kinetic Model Incorporated Loss for Denoising Arterial Spin Labeled MRI via Deep Residual Learning

Keyword [Cerebral Blood Flow]

Ulas C, Tetteh G, Kaczmarz S, et al. DeepASL: Kinetic Model Incorporated Loss for Denoising Arterial Spin Labeled MRI via Deep Residual Learning[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2018: 30-38.



1. Overview


1.1. Motivation

Arterial spin labeling (ASL) allows to quantify the cerebral blood flow (CBF) by magnetic labeling of the arterial blood water, but suffers from an inherently low-signal-to-noise ratio (SNR).

In this paper

  • FCN to learn residual from noisy perfusion-weighted image
  • incorporate the CBF estimation model in the loss function during training

1.2. Model



  • Input. 2D noisy gray image patches
  • 8 Conv2D (48 channels, 3x3, following pReLU) + 1 Conv2D (no activation)
  • training. 18000 patch pairs of 40x40, batch size 500, 200 epoch, LR 0.0001

1.3. CBF Estimation Model



  • β. brain-blood partition coefficient
  • T_{1b}. longitudinal relaxation time of blood
  • α. labeling efficiency
  • τ. label duration
  • PLD. post-label delay
  • SI_{PD}. proton density weighted image
  • ΔM. perfusion-weighted images

1.4. Loss Function



  • N. noisy
  • f_t. reference CBF value for each voxel

1.5. Experiments