Synthetic Data for Robust Stroke Segmentation

2024-04-02

arXiv.org

AI Summary

Stroke lesion segmentation typically relies on high-resolution images and large-scale annotated datasets, resulting in poor clinical generalizability. Method: We propose the first synthetic-data framework tailored for stroke pathophysiology modeling. Building upon SynthSeg, we introduce a lesion-driven synthetic augmentation strategy that generates multi-sequence MRI data featuring diverse stroke characteristics—including infarction, hemorrhage, and edema. We further design a universal segmentation model that jointly models healthy and pathological tissues without sequence-specific training, implemented via an enhanced nnUNet architecture integrated with SPM-derived anatomical priors and optimized PyTorch execution. Contribution/Results: Our method achieves state-of-the-art performance on in-domain data and significantly outperforms existing approaches on out-of-domain data across multiple clinical centers and scanner types. It substantially reduces dependency on real annotated data while markedly improving clinical robustness and generalizability.

Technology Category

• Computer Vision: Segmentation
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Neuro-Symbolic Learning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Current deep learning-based approaches to lesion segmentation in neuroimaging often depend on high-resolution images and extensive annotated data, limiting clinical applicability. This paper introduces a novel synthetic data framework tailored for stroke lesion segmentation, expanding the SynthSeg methodology to incorporate lesion-specific augmentations that simulate diverse pathological features. Using a modified nnUNet architecture, our approach trains models with label maps from healthy and stroke datasets, facilitating segmentation across both normal and pathological tissue without reliance on specific sequence-based training. Evaluation across in-domain and out-of-domain (OOD) datasets reveals that our method matches state-of-the-art performance within the training domain and significantly outperforms existing methods on OOD data. By minimizing dependence on large annotated datasets and allowing for cross-sequence applicability, our framework holds potential to improve clinical neuroimaging workflows, particularly in stroke pathology. PyTorch training code and weights are publicly available at https://github.com/liamchalcroft/SynthStroke, along with an SPM toolbox featuring a plug-and-play model at https://github.com/liamchalcroft/SynthStrokeSPM.

Problem

• Overcoming reliance on high-resolution images and annotated data for stroke lesion segmentation
• Enhancing stroke lesion segmentation with synthetic data and diverse pathological simulations
• Improving cross-sequence applicability and clinical neuroimaging workflows in stroke pathology

Innovation

• Synthetic data framework for stroke segmentation
• Modified nnUNet with lesion-specific augmentations
• Cross-sequence applicability without sequence-based training