SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining (2107.09559v4)

Abstract: Despite advances in data augmentation and transfer learning, convolutional neural networks (CNNs) difficultly generalise to unseen domains. When segmenting brain scans, CNNs are highly sensitive to changes in resolution and contrast: even within the same MRI modality, performance can decrease across datasets. Here we introduce SynthSeg, the first segmentation CNN robust against changes in contrast and resolution. SynthSeg is trained with synthetic data sampled from a generative model conditioned on segmentations. Crucially, we adopt a domain randomisation strategy where we fully randomise the contrast and resolution of the synthetic training data. Consequently, SynthSeg can segment real scans from a wide range of target domains without retraining or fine-tuning, which enables straightforward analysis of huge amounts of heterogeneous clinical data. Because SynthSeg only requires segmentations to be trained (no images), it can learn from labels obtained by automated methods on diverse populations (e.g., ageing and diseased), thus achieving robustness to a wide range of morphological variability. We demonstrate SynthSeg on 5,000 scans of six modalities (including CT) and ten resolutions, where it exhibits unparalleled generalisation compared with supervised CNNs, state-of-the-art domain adaptation, and Bayesian segmentation. Finally, we demonstrate the generalisability of SynthSeg by applying it to cardiac MRI and CT scans.

An Analysis of SynthSeg's Contribution to Robust MRI Segmentation

The paper "SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining" outlines a novel approach to brain MRI segmentation using convolutional neural networks (CNNs). SynthSeg introduces a strategy that enhances the generalization abilities of CNNs across various domain shifts, specifically changes in contrast and resolution, without the need for retraining the model for different target domains. This examination scrutinizes the methods, results, and implications for both practice and further research.

SynthSeg addresses a fundamental shortcoming in current CNN-based segmentation techniques: the limited capability to generalize across different imaging conditions such as contrast variations and resolutions without domain-specific retraining. Traditional CNNs, including state-of-the-art models like nnUNet, typically degrade in performance when applied to conditions outside their trained domain. SynthSeg counters this limitation by employing synthetic data generated through a process of domain randomization. This synthetic data is created on-the-fly through a generative model inspired by Bayesian segmentation, where the model generates synthetic images with random contrasts, resolutions, and other variations, thus encompassing a wide array of imaging conditions.

Key results from the application of SynthSeg demonstrate its robustness across a spectrum of MRI modalities and resolutions, including T1, T2, FLAIR, and even CT images. The model’s performance was compared against several methods: the supervised T1 baseline, nnUNet, domain adaptation strategies such as TTA and SIFA, and Bayesian segmentation with SAMSEG. SynthSeg consistently achieved superior Dice scores across all tests, notably outperforming SAMSEG on low-resolution scans, which are especially challenging due to partial volume effects not considered by typical Bayesian methods.

One of SynthSeg’s critical innovations is its ability to maintain robustness through training with only a set of anatomical label maps, thus requiring no real images for training. This approach significantly alleviates the need for extensive and costly manual labelling, often a bottleneck in medical dataset preparation. Furthermore, the findings indicate that utilizing synthetic scans generated with highly randomized parameters helps the model learn representations that are agnostic to specific domain characteristics, enhancing cross-domain applicability and reducing the necessity for domain-specific adaptation procedures.

A substantial implication of SynthSeg's technology is its potential for broader clinical application, where MRI scans vary widely in terms of acquisition protocols. By eliminating the need for retraining or fine-tuning for each particular domain, SynthSeg simplifies the deployment of robust MRI segmentation solutions in heterogeneous clinical environments. The paper supports this claim through experimentation across diverse datasets, demonstrating the accuracy and reliability of SynthSeg when assessing conditions like Alzheimer's Disease through volumetric studies despite variations in image acquisition.

In applying the model to cardiac MRI and CT images, SynthSeg further evidenced its versatility. The model was trained on limited, primarily synthetic, cardiac datasets, yet still delivered results competitive with those in the existing literature requiring extensive retraining. This application exemplifies SynthSeg's potential beyond neuroimaging, hinting at its usability for wider biomedical imaging challenges.

The paper establishes a strong precedent for using domain randomization in medical image segmentation, potentially guiding future research towards leveraging synthetically generated training data to overcome domain generalization barriers. However, the paper's reliance on synthetic data necessitates careful design of the generative model to ensure robustness across real-world data, underscoring an area for ongoing refinement.

In conclusion, SynthSeg represents a significant step forward in the field of medical image analysis, offering a strategy that generalizes well across different imaging domains without additional retraining costs. This work not only holds promise for practical adoption in diverse clinical settings but also provides a conceptual framework for future advancements in domain-agnostic segmentation models. Further exploration of SynthSeg's applications beyond brain MRI, its integration with other medical imaging modalities, and its combination with other generative techniques could provide valuable directions for subsequent research in the area of automated medical image segmentation.