[Paper Review] A Learning Strategy for Contrast-agnostic MRI Segmentation

Benjamin Billot, Douglas N. Greve+4 more|arXiv (Cornell University)|Mar 4, 2020

Medical Image Segmentation Techniques19 references47 citations

TL;DR

SynthSeg trains a CNN using synthetic, contrast-varied scans generated from a Bayesian segmentation model, enabling contrast-agnostic brain MRI segmentation without re-training for new modalities. It achieves competitive accuracy with much faster inference and strong cross-dataset generalization.

ABSTRACT

We present a deep learning strategy that enables, for the first time, contrast-agnostic semantic segmentation of completely unpreprocessed brain MRI scans, without requiring additional training or fine-tuning for new modalities. Classical Bayesian methods address this segmentation problem with unsupervised intensity models, but require significant computational resources. In contrast, learning-based methods can be fast at test time, but are sensitive to the data available at training. Our proposed learning method, SynthSeg, leverages a set of training segmentations (no intensity images required) to generate synthetic sample images of widely varying contrasts on the fly during training. These samples are produced using the generative model of the classical Bayesian segmentation framework, with randomly sampled parameters for appearance, deformation, noise, and bias field. Because each mini-batch has a different synthetic contrast, the final network is not biased towards any MRI contrast. We comprehensively evaluate our approach on four datasets comprising over 1,000 subjects and four types of MR contrast. The results show that our approach successfully segments every contrast in the data, performing slightly better than classical Bayesian segmentation, and three orders of magnitude faster. Moreover, even within the same type of MRI contrast, our strategy generalizes significantly better across datasets, compared to training using real images. Finally, we find that synthesizing a broad range of contrasts, even if unrealistic, increases the generalization of the neural network. Our code and model are open source at https://github.com/BBillot/SynthSeg.

Motivation & Objective

• Motivate robust brain MRI segmentation across arbitrary contrasts without additional training for new modalities.
• Leverage a generative Bayesian segmentation model to synthesize diverse training data.
• Train a CNN on segmentation maps with on-the-fly synthetic contrasts to achieve contrast-agnostic performance.

Proposed method

• Use the classical generative model for Bayesian MRI segmentation as the backbone.
• Condition a CNN on segmentation maps to generate synthetic scans with random appearance, deformation, noise, and bias field parameters.
• Via mini-batch sampling of randomly varied model parameters, produce diverse synthetic contrasts to prevent contrast bias in the learned features.
• Train to perform segmentation directly on unpreprocessed brain MRI scans across various contrasts.
• Evaluate on multiple datasets with four MRI contrasts to assess generalization and speed.
• Demonstrate that synthesis-based training rivals Bayesian segmentation in accuracy while being three orders of magnitude faster.

Experimental results

Research questions

• RQ1 Can a CNN be trained to segment brain MRIs across unseen contrasts without modality-specific paired data or fine-tuning?
• RQ2 Does synthetic contrast augmentation based on a Bayesian generative model enable better cross-dataset generalization than training on real images?
• RQ3 What is the impact of synthesizing a broad and sometimes unrealistic range of contrasts on network generalization?

Key findings

• The approach segments every contrast in the evaluated data.
• SynthSeg is slightly faster and competitive with classical Bayesian segmentation in accuracy.
• Generalization across datasets within the same MRI contrast is significantly better than CNNs trained on real images.
• Training with a broad range of, even unrealistic, synthesized contrasts improves generalization.
• The method scales to four datasets and over 1,000 subjects across four MR contrasts.