Liam Chalcroft

Liam Chalcroft

ML PhD @ UCL | Deep Learning and 3D Computer Vision for Medical Imaging

Synthetic Data for Robust Stroke Segmentation

Published in arXiv, 2024

Recommended citation: Chalcroft, L., Pappas, I., Price, C.J., & Ashburner, J. (2024). Synthetic Data for Robust Stroke Segmentation. arXiv preprint. arXiv:2404.01946v1. http://arxiv.org/abs/2404.01946v1

Abstract

Deep learning-based semantic segmentation in neuroimaging currently requires high-resolution scans and extensive annotated datasets, posing significant barriers to clinical applicability. We present a novel synthetic framework for the task of lesion segmentation, extending the capabilities of the established SynthSeg approach to accommodate large heterogeneous pathologies with lesion-specific augmentation strategies. Our method trains deep learning models, demonstrated here with the UNet architecture, using label maps derived from healthy and stroke datasets, facilitating the segmentation of both healthy tissue and patho- logical lesions without sequence-specific training data. Evaluated against in-domain and out-of-domain (OOD) datasets, our framework demon- strates robust performance, rivaling current methods within the training domain and significantly outperforming them on OOD data. This contri- bution holds promise for advancing medical imaging analysis in clinical settings, especially for stroke pathology, by enabling reliable segmenta- tion across varied imaging sequences with reduced dependency on large annotated corpora.

Code: SynthStroke

Key Figures

Sample predictions with ensembling method

Figure 3: Sample predictions with ensembling method in the ISLES 2015 dataset. Color code in predicted Synth labels: dark-blue=GM, light-blue=WM, green=PV, amber=CSF, red=stroke.

Summary

This paper presents a novel synthetic framework for robust stroke lesion segmentation, extending the SynthSeg approach to accommodate large heterogeneous pathologies. The key contributions and findings include:

  1. A method that combines synthetic data generation with real data from the ATLAS dataset to train deep learning models for stroke lesion segmentation.

  2. The use of lesion-specific augmentation strategies, including simulating varying grades of infarction and stitching lesions into surrounding tissue structures.

  3. Evaluation on both in-domain (ATLAS) and out-of-domain (ISLES 2015 and ISLES 2022) datasets, demonstrating the model’s robustness and generalizability.

  4. Competitive performance with baseline models on in-domain data and significant improvements on out-of-domain data.

  5. The ability to perform multi-modal ensembling, which outperforms single-modality predictions and reduces the likelihood of mislabeling irrelevant hyperintensities.

  6. Potential applicability to other similar domains such as hemorrhage and glioblastoma.

The study demonstrates that it is possible to augment a SynthSeg-style training procedure to include large, heterogeneous lesions, potentially advancing medical imaging analysis in clinical settings by enabling reliable segmentation across varied imaging sequences with reduced dependency on large annotated datasets.

Download paper here