Source-linked AI summary
Learning Implicit Brain MRI Manifolds with Deep Learning
Camilo Bermudez, Andrew J. Plassard, Larry T. Davis, Allen T. Newton, Susan M Resnick, Bennett A. Landman
TL;DR
Previous brain-MRI manifold mappings have been limited by explicit similarity measures, and prior synthesis work lacked image-quality validation by imaging experts. This paper applies deep learning to normal-brain MRI synthesis and denoising; synthetic-image quality scores substantially overlap real-image scores, while denoising has higher PSNR than FSL SUSAN.
Problem
Brain-MRI manifold approaches have relied on explicit similarity measures, while earlier near-realistic synthesis results had not been validated for image quality by imaging experts.
Source: §1 · pp. 1–2Method
Deep learning is used to explore the implicit manifold of normal brains through GAN-based synthesis and skip-connected autoencoder denoising.
Source: §1 · pp. 1–2Results
Synthetic and real image quality-score distributions substantially overlap, and the proposed denoising method has greater PSNR than FSL SUSAN at all three noise levels (p < 0.001).
Source: §3.1 · p. 4, §3.2 · p. 5Takeaways & Limitations
The paired tasks provide a way to explore an implicit brain-MRI manifold and generate realistic images spanning the normal-brain domain.
Source: §1 · p. 2Synthetic images could receive comparable quality ratings yet show anatomical abnormalities, including asymmetric caudate size, and unrealistic images may not be anatomically viable.
Source: §4 · p. 5
Abstract
from arXiv · showhide
An important task in image processing and neuroimaging is to extract quantitative information from the acquired images in order to make observations about the presence of disease or markers of development in populations. Having a lowdimensional manifold of an image allows for easier statistical comparisons between groups and the synthesis of group representatives. Previous studies have sought to identify the best mapping of brain MRI to a low-dimensional manifold, but have been limited by assumptions of explicit similarity measures. In this work, we use deep learning techniques to investigate implicit manifolds of normal brains and generate new, high-quality images. We explore implicit manifolds by addressing the problems of image synthesis and image denoising as important tools in manifold learning. First, we propose the unsupervised synthesis of T1-weighted brain MRI using a Generative Adversarial Network (GAN) by learning from 528 examples of 2D axial slices of brain MRI. Synthesized images were first shown to be unique by performing a crosscorrelation with the training set. Real and synthesized images were then assessed in a blinded manner by two imaging experts providing an image quality score of 1-5. The quality score of the synthetic image showed substantial overlap with that of the real images. Moreover, we use an autoencoder with skip connections for image denoising, showing that the proposed method results in higher PSNR than FSL SUSAN after denoising. This work shows the power of artificial networks to synthesize realistic imaging data, which can be used to improve image processing techniques and provide a quantitative framework to structural changes in the brain.
1. INTRODUCTION
Brain MRI manifold methods have been limited by explicit similarity requirements, incomplete representations, or overly smoothed interpolations. This work proposes deep-learning synthesis and denoising as tools for exploring an implicit manifold of normal brains.
1. INTRODUCTION
- Earlier brain representations include atlases, linear mappings, Isomap, and ROI-based regressions, but the approaches have stated limitations.ROI measures do not comprehensively map subject space, linear representations miss nonlinear effects such as aging, and Isomap requires explicit similarity measures and interpolates potentially over-smoothed images.Source: §1 · p. 1
- Deep autoencoders and convolutional neural networks offer nonlinear functions for learning manifold representations, motivating the paper's synthesis and denoising tasks.The cited approaches can be supervised or unsupervised.Source: §1 · p. 1
- Prior synthesis work focused on resolution enhancement and inter-modality generation, and its near-realistic images had not been validated for quality by imaging experts.The paper also frames denoising as improving image-to-manifold mapping and the ability of manifold distances to distinguish brain representations.Source: §1 · p. 2
- The paper uses GANs for unsupervised brain-MRI synthesis and skip-connected autoencoders for denoising, reporting better denoising PSNR than FSL SUSAN.Together, the tools are proposed as a way to map an implicit brain-MRI manifold and generate realistic images spanning the normal-brain domain.Source: §1 · p. 2
2. METHODS
The methods use healthy-control T1-weighted MRI slices to train a GAN for image synthesis and a skip-connected autoencoder for denoising, with separate procedures for evaluating each task.
2.1 Dataset
- 528 healthy-control T1-weighted MRI scans supplied one midline axial slice each, affine-registered to MNI space and intensity-normalized.The participants were ages 29–94 years (mean 67.9); slices were 220 × 172 voxels, acquired at 1 mm isotropic resolution.Source: §2.1 · p. 2
2.2 Generative Adversarial Networks for Image Synthesis
- The GAN generator upsamples uniform-noise vectors through 2D deconvolutional layers, while the discriminator classifies real and synthesized images.The discriminator uses four 2D convolutional layers, LeakyReLU and dropout, then a sigmoid classifier; the coupled networks train the generator to produce more realistic images.Source: Figure 1 · p. 2, §2.2 · p. 3
- The GAN trained for 1,500 epochs on 528 noise vectors, with 100 new noise vectors for testing and image correlation used to check for training-image replication.Two raters—a neuroradiologist and an imaging expert—scored a random set of 100 real and 100 synthesized images on a 1–5 quality scale.Source: §2.2 · p. 3
2.3 Denoising Autoencoder
- For denoising, three noisy versions per subject were generated at SNR 1, 10, and 100, and separate autoencoders encoded each into a high-dimensional representation before reconstruction.The network adds convolutional features to corresponding deconvolutional layers to preserve structural features while minimizing mean squared error against the original image.Source: §2.3 · pp. 3–4, Figure 2 · p. 3
- Denoising was evaluated against FSL SUSAN v5.0 using PSNR, with SUSAN parameters optimized on SNR 10 images and reused at SNR 1 and 100.The comparison measured PSNR for the noisy input, SUSAN output, and proposed method.Source: §2.3 · p. 4
3. RESULTS
The synthesis network produced images with structural differences from highly correlated real scans, and expert quality scores overlapped substantially. The denoising method achieved higher PSNR than FSL SUSAN across noise levels.
3.1 Brain MRI Synthesis
- Highly correlated real images differed from synthesized images in features such as ventricle size and gyral patterns, suggesting the network did not simply reproduce training examples.The three real-image examples had correlations of 0.76, 0.77, and 0.78 with a synthesized image.Source: §3.1 · p. 4, Figure 3 · p. 4
- Expert quality-score distributions substantially overlapped, although the reported average scores were 4.9 for real images and 3.9 for synthetic images.The histograms show scores from a neuroradiologist and an imaging expert.Source: §3.1 · p. 4, Figure 4 · p. 5
3.2 Image Denoising
- Across all three noise levels, the proposed denoising method produced greater PSNR than FSL SUSAN, with the difference significant by Wilcox rank sum test (p < 0.001).Source: §3.2 · p. 5
4. DISCUSSION & CONCLUSION
The discussion reports promising synthesis and denoising results while identifying anatomical accuracy, signal quality, and broader data coverage as constraints. Denoising fidelity was recovered at higher SNR, but smoothing remained at SNR 1.
4. DISCUSSION & CONCLUSION
- Synthetic images had anatomical and signal-quality issues, including caudate asymmetry and brighter central intensities; structural restrictions and intensity constraints were proposed as possible refinements.The discussion notes that unrealistic synthesized images may not represent anatomically viable MRI images.Source: §4 · p. 5
- Extending denoising to 3D brain volumes remains future work and may face resource challenges from network depth and problem dimensionality.Source: §4 · p. 5
- Promising results from 528 images were likely aided by a relatively homogeneous training set and a simple task; broader 3D data across ages, centers, and acquisition parameters were proposed for improved fidelity.The discussion also notes that contextual features could further increase image fidelity.Source: §4 · pp. 5–6
- The denoising method showed higher reconstructed PSNR across noise levels, while fidelity recovery was reported at higher SNR and smoothing at SNR 1.Figure 5 presents noisy inputs at SNR 1, 10, and 100 alongside denoised outputs and PSNR comparisons.Source: §4 · p. 5, Figure 5 · p. 6