Transfer Learning Promotes Robust Parametric Mapping of Diffusion Encoded MR FingerprintingDownload PDF

Published: 09 May 2022, Last Modified: 12 May 2023MIDL 2022 Short PapersReaders: Everyone
Keywords: Transfer Learning, MR Fingerprinting, Recurrent Neural Network, Diffusion
TL;DR: Transfer learning model resulted in more reliable reconstructions for T1, T2, and ADC for a diffusion encoded MR Fingerprinting sequence in the setting of noise.
Abstract: MR fingerprinting (MRF) is a framework to simultaneously quantify multiple tissue properties. Acquisition parameters are varied pseudo-randomly and each signal evolution is matched with a dictionary of simulated entries. However, dictionary methods are computationally and memory intensive. Deep learners (DL) are capable of mapping complex MRF signal evolutions to a quantitative parametric space, reducing the computational requirements and reconstruction time; yet fail to perform as well in the setting of noise. Drawing from natural language processing (NLP) we proposed a transfer learning (TL) model to improve MRF parametric estimates with realistic noise levels. The weights of a network trained on clean data are used to instantiate the weights of a noisy model. The model is constrained to learn noise invariant features, by freezing the last layer. Signal evolutions were modeled using a recurrent neural network (RNN) to reconstruct T1, T2, and the apparent diffusion coefficient (ADC). Compared to a model trained with noise, but without TL our approached resulted in a 15% reduction in mean squared error (MSE). Monte Carlo simulations performed at varying SNR (10-60 dB) showed our method yielded losses comparable to the clean model at higher SNRs and proved more robust in the setting of noise at lower SNRs.
Registration: I acknowledge that acceptance of this work at MIDL requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: novel methodological ideas without extensive validation
Primary Subject Area: Learning with Noisy Labels and Limited Data
Secondary Subject Area: Transfer Learning and Domain Adaptation
Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.
1 Reply