Skull stripping with purely synthetic data

TL;DR

This paper tackles skull stripping across modalities and species without relying on real brain images or labels by introducing PUMBA, a purely synthetic approach. It generates ellipsoid-based synthetic head phantoms with random deformations and intensities to train a 3D U-Net in a fully unsupervised fashion, aiming for broad generalization. The model achieves competitive accuracy on multi-modal human data (TCGA, IXI, LPBA40, MINDS) and shows strong performance on non-human primate and rodent datasets, including a high accuracy in marmosets, without image priors. The work suggests a generalizable direction for medical image segmentation when ground-truth data are scarce, and provides code for replication.

Abstract

While many skull stripping algorithms have been developed for multi-modal and multi-species cases, there is still a lack of a fundamentally generalizable approach. We present PUMBA(PUrely synthetic Multimodal/species invariant Brain extrAction), a strategy to train a model for brain extraction with no real brain images or labels. Our results show that even without any real images or anatomical priors, the model achieves comparable accuracy in multi-modal, multi-species and pathological cases. This work presents a new direction of research for any generalizable medical image segmentation task.

Figures (6)

• Figure 1: A visual example of how the synthetic training data is generated. The steps of generating ellipsoids and applying random intensities and deformations is shown on a. Example images from the training dataset is shown on b. Green denotes the main mask, while blue denotes the boundaries.
• Figure 2: The effect of the post processing step is shown on a T1 weighted image from the IXI dataset. Green (main mask) and blue region (boundary) is predicted from the model, which is processed to create the sole green mask.
• Figure 3: The quantitative metrics are shown on the TCGA (top row), MINDS (middle row) and LPBA40 (bottom row) dataset. Dice Score (first column), Jaccard Index (second column) and Hausdorff Distance (third column) is shown. Note that unlike Dice Score and Jaccard Index, accurate segmentation results in smaller Hausdorff Distance.
• Figure 4: The predicted mask of each method on multiple modalities of the TCGA (tumor) and the IXI (healthy) dataset. Ground truth is not provided for the IXI dataset.
• Figure 5: The segmentation results on the MINDS (Marmoset) dataset and CAMRI (Rodent) dataset. Ground truth is shown where applicable.