Efficient Brain Extraction of MRI Scans with Mild to Moderate Neuropathology

Skull stripping magnetic resonance images (MRI) of the human brain is an important process in many image processing techniques, such as automatic segmentation of brain structures. Numerous methods have been developed to perform this task, however, they often fail in the presence of neuropathology and can be inconsistent in defining the boundary of the brain mask. Here, we propose a novel approach to skull strip T1-weighted images in a robust and efficient manner, aiming to consistently segment the outer surface of the brain, including the sulcal cerebrospinal fluid (CSF), while excluding the full extent of the subarachnoid space and meninges. We train a modified version of the U-net on silver-standard ground truth data using a novel loss function based on the signed-distance transform (SDT). We validate our model both qualitatively and quantitatively using held-out data from the training dataset, as well as an independent external dataset. The brain masks used for evaluation partially or fully include the subarachnoid space, which may introduce bias into the comparison; nonetheless, our model demonstrates strong performance on the held-out test data, achieving a consistent mean Dice similarity coefficient (DSC) of 0.964$\pm$0.006 and an average symmetric surface distance (ASSD) of 1.4mm$\pm$0.2mm. Performance on the external dataset is comparable, with a DSC of 0.958$\pm$0.006 and an ASSD of 1.7$\pm$0.2mm. Our method achieves performance comparable to or better than existing state-of-the-art methods for brain extraction, particularly in its highly consistent preservation of the brain’s outer surface. The method is publicly available on GitHub.

💡 Research Summary

The paper introduces a novel deep‑learning approach for skull stripping of T1‑weighted brain MRI scans that contain mild to moderate neuropathology. Traditional skull‑stripping tools such as BET, MONSTR, ROBEX, and Synthstrip either struggle with pathological brains, require long processing times, or depend heavily on high‑quality manual annotations. To address these issues, the authors propose a modified 3D U‑Net trained on “silver‑standard” masks—probabilistic ground‑truth generated by fusing the outputs of MONSTR, ROBEX, and Synthstrip using the STAPLE algorithm. This strategy circumvents the need for labor‑intensive manual labeling while still capturing consensus information from several state‑of‑the‑art methods.

A key technical contribution is the loss function based on the signed‑distance transform (SDT). The SDT assigns zero to voxels on the mask boundary, positive values inside the brain, and negative values outside, thus providing a continuous distance field. The authors formulate a weighted mean‑squared‑error loss (L_SDT) where each voxel weight follows an exponential decay w_i = exp(‑α|ŷ_i|) + β. Hyper‑parameters α = 0.1 and β = 0.3 were tuned with Optuna. This weighting concentrates gradients near the boundary, promotes smoother convergence, and makes training robust to imperfections in the silver‑standard masks—an advantage over the commonly used Dice loss, which performed sub‑optimally on these noisy labels.

The network architecture is deliberately lightweight: depth = 3, initial channels = 32, with transposed convolutions for up‑sampling. All images are rigidly registered to MNI space, intensity‑normalized, and down‑sampled by a factor of two to reduce memory consumption, enabling larger batch sizes and a total training time of less than 12 hours on two RTX 4000 GPUs. During inference, the predicted continuous SDT map is thresholded at zero to obtain a binary brain mask; disconnected components are removed and small holes are filled as post‑processing.

Evaluation was performed on three fronts. First, a held‑out test set comprising 547 scans from ADNI and ASAP (with silver‑standard masks) yielded a mean Dice similarity coefficient (DSC) of 0.964 ± 0.006, average symmetric surface distance (ASSD) of 1.4 ± 0.2 mm, and 95th percentile Hausdorff distance (HD95) of 3.6 ± 1.1 mm. These results are on par with or slightly better than ROBEX (DSC = 0.969) and MONSTR (DSC = 0.964) while offering more consistent outer‑surface preservation.

Second, the model was tested on external, unseen datasets: IXI (n = 50), FSM (n = 38), and an ASL cohort (n = 43). Across these, the method achieved DSCs of 0.958 ± 0.006, 0.956 ± 0.008, and 0.958 ± 0.007 respectively, with ASSDs ranging from 1.7 to 1.9 mm. Notably, the reference masks in these datasets include subarachnoid CSF, whereas the proposed method deliberately excludes it; despite this discrepancy, performance remains comparable, highlighting the model’s ability to generalize across differing annotation conventions.

Third, a small manually labeled set (CC359, n = 12) was used to assess performance against high‑quality human annotations that exclude sulcal CSF. Here the method obtained DSC = 0.945 ± 0.005, ASSD = 3.1 ± 0.6 mm, and HD95 = 11.5 ± 3.3 mm, outperforming ROBEX and Synthstrip on this stringent benchmark.

Qualitative inspection confirmed that the proposed network consistently preserves the brain’s outer surface, avoids over‑segmentation of meninges, and remains robust to motion artefacts and pathological deformations. Failure cases (≈0.7 % of images) involved occasional over‑segmentation into high‑intensity neck fat, but the boundary remained accurate. Inference time averages under 20 seconds per scan, comparable to other deep‑learning skull‑stripping tools.

The authors conclude that the combination of down‑sampled training, the SDT‑based loss, and silver‑standard supervision yields a fast, accurate, and pathology‑robust brain extraction method. Limitations include reliance on consensus masks that may still contain systematic biases and the potential loss of fine cortical detail due to down‑sampling. Future work will explore multi‑resolution architectures, richer post‑processing (e.g., CRF), and broader training data encompassing more disease phenotypes. The full code and pretrained model are publicly released on GitHub, facilitating reproducibility and adoption by the neuroimaging community.