In this project I designed a deep learning network to accelerate 4D data reconstruction in Tensorflow.
In this work 4D Angiography and Perfusion at eight time-points are reconstructed from an interleaved half-sampled crushed and non-crushed Hadamard-te arterial spin labeling (ASL) of rank 8. The network uses DenseUnet structure and multi-stage loss function. Different loss functions have been applied for training including: perceptual loss (PL), mean squre error (MSE), Structural Similarity Index (SSIM) in a single and multi-stage fasions. Also, a framework for generating dynamic ASL scans based on the Hadamard ASL kinetic model has been proposed.
The reconstruction process can be formulated as:
,
in which
is the decoding and subtraction function,
and
are the acquired scans of the
row of non-crushed and crushed Hadamard te-pCASL datasets,
and
denote perfusion and angiography scans respectively.
Here you can find the Hadamard te-ASL signal generator.
@inproceedings{yousefi2019fast,
title={Fast Dynamic Perfusion and Angiography Reconstruction using an end-to-end 3D Convolutional Neural Network},
author={Yousefi, Sahar and Hirschler, Lydiane and van der Plas, Merlijn and Elmahdy, Mohamed S and Sokooti, Hessam and Van Osch, Matthias and Staring, Marius},
booktitle={International Workshop on Machine Learning for Medical Image Reconstruction},
pages={25--35},
year={2019},
organization={Springer}
}
Figure 1- Proposed network, a multi-stage DenseUnet. Inputs: an interleaved half-sampled crushed and non-crushed Hadamard-te arterial spin labeling (ASL) of rank 8. Output: dynamic angiography and perdusion scans at 8 time-point.
Figure 2- Proposed data generator.
Figure 3- Results of reconstructed angiography scans for one subject.
Figure 4- Results of reconstructed perfusion scans for one subject.
Tensorflow<2 & python>3.4