[Paper Review] Deep learning brain conductivity mapping using a patch-based 3D U-net

Nils Hampe, Ulrich Katscher+3 more|arXiv (Cornell University)|Aug 12, 2019

Electrical and Bioimpedance Tomography21 references4 citations

TL;DR

This study proposes a patch-based 3D U-Net deep learning model to reconstruct quantitative brain conductivity maps from B1 transceive phase data using electrical properties tomography (EPT). Trained on simulated data with added Gaussian noise, the network produces high-quality reconstructions, but artifacts emerge when applied to in-vivo data; however, fine-tuning on in-vivo phase data with conventional EPT-derived conductivity labels significantly reduces artifacts, demonstrating improved robustness to real-world acquisition noise and systematic discrepancies.

ABSTRACT

Purpose: To investigate deep learning electrical properties tomography (EPT) for application on different simulated and in-vivo datasets including pathologies for obtaining quantitative brain conductivity maps. Methods: 3D patch-based convolutional neural networks were trained to predict conductivity maps from B1 transceive phase data. To compare the performance of DLEPT networks on different datasets, three datasets were used throughout this work, one from simulations and two from in-vivo measurements from healthy volunteers and cancer patients, respectively. At first, networks trained on simulations are tested on all datasets with different levels of homogeneous Gaussian noise introduced in training and testing. Secondly, to investigate potential robustness towards systematical differences between simulated and measured phase maps, in-vivo data with conductivity labels from conventional EPT is used for training. Results: High quality of reconstructions from networks trained on simulations with and without noise confirms the potential of deep learning for EPT. However, artifact encumbered results in this work uncover challenges in application of DLEPT to in-vivo data. Training DLEPT networks on conductivity labels from conventional EPT improves quality of results. This is argued to be caused by robustness to artifacts from image acquisition. Conclusions: Networks trained on simulations with added homogeneous Gaussian noise yield reconstruction artifacts when applied to in-vivo data. Training with realistic phase data and conductivity labels from conventional EPT allows for severely reducing these artifacts.

Motivation & Objective

• To evaluate deep learning-based EPT (DLEPT) for generating quantitative brain conductivity maps across diverse datasets including simulations and in-vivo scans.
• To assess the robustness of DLEPT networks trained on simulated data when applied to real in-vivo data with inherent artifacts and noise.
• To investigate whether training on in-vivo phase data with ground-truth conductivity labels from conventional EPT improves reconstruction quality and artifact reduction.
• To compare the performance of DLEPT models trained on simulations versus real in-vivo data across healthy and pathological brain tissues.
• To determine the impact of noise and systematic differences between simulated and measured phase maps on DLEPT reconstruction fidelity.

Proposed method

• A 3D patch-based U-Net architecture is employed to predict conductivity maps from B1 transceive phase data, enabling local spatial modeling and efficient processing of volumetric MRI data.
• The network is first trained on simulated datasets with varying levels of homogeneous Gaussian noise to assess generalization and robustness.
• Subsequently, the model is fine-tuned using in-vivo phase data from healthy volunteers and cancer patients, paired with conductivity labels generated via conventional EPT.
• Data augmentation and normalization techniques are applied to ensure consistency across diverse input distributions from simulations and real scans.
• Performance is evaluated using quantitative metrics such as mean squared error and structural similarity index (SSIM) across different datasets.
• The model leverages skip connections and 3D convolutions to preserve spatial resolution and capture hierarchical features in 3D brain tissue structures.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on simulated EPT data generalize to in-vivo brain data without significant degradation in reconstruction quality?
• RQ2 How do artifacts in in-vivo data affect the performance of DLEPT networks trained exclusively on simulated data?
• RQ3 Does fine-tuning a DLEPT model on in-vivo phase data with conventional EPT-derived conductivity labels improve reconstruction accuracy and reduce artifacts?
• RQ4 To what extent does training on realistic phase data mitigate the impact of systematic differences between simulated and measured B1 phase maps?
• RQ5 How does the addition of Gaussian noise during training affect the robustness of DLEPT models when deployed on real in-vivo datasets?

Key findings

• Networks trained on simulated data with added Gaussian noise produced high-quality conductivity reconstructions on simulated data but exhibited significant artifacts when applied to in-vivo datasets.
• Reconstructions from models trained solely on simulations showed reduced fidelity on in-vivo data, indicating poor generalization due to discrepancies between simulated and real-phase maps.
• Fine-tuning the DLEPT model on in-vivo phase data with conductivity labels from conventional EPT significantly reduced artifacts and improved reconstruction quality.
• The improved performance after in-vivo fine-tuning was attributed to better robustness against real-world acquisition artifacts and systematic phase map distortions.
• The results suggest that training on realistic in-vivo data is essential for reliable DLEPT application in clinical neuroimaging settings.
• The study confirms that data domain shift between simulation and in-vivo data is a major challenge for DLEPT, and that domain-specific fine-tuning is necessary for clinical viability.