Introduction

Brain extraction is a preliminary but critical step in many neuroimaging studies and determines the accuracy of subsequent analyses. Standard brain extraction algorithms are, however, limited to the processing of precontrast T1-weighted (T1-w) MRI and frequently fail in the presence of pathologically altered brain. To overcome these limitations we utilize MRI data from a large multicenter clinical trial in neuro-oncology to develop, train and independently validate an ANN for brain extraction. Specifically, we aimed to develop an automated method that (a) acts robust in the presence of pathology or treatment-induced tissue alterations, varying MRI hardware and acquisition parameters, (b) does not require any pre-processing, and (c) is applicable to independently process any type of anatomical MRI sequence.

Methods

We used MRI data of patients with recurrent glioblastoma from a prospective randomized phase II and III trial (EORTC 26101 study) with 2495 examinations from 590 patients acquired across 37 centers within the Europe. Data included pre- and postcontrast T1-weighted (T1-w, cT1-w), FLAIR and T2-weighted (T2-w) images and ground-truth brain mask segmentations. The ANN was trained and validated on 2/3 of MRI-data from EORTC 26101 study and tested on the remaining 1/3 of MRI. Additionally, for independent evaluation the ANN was tested on three public available datasets (LPBA40, NFBS, CC-359). We evaluated the performance using DICE coefficient and Hausdorff distance in comparison with six standard BE-algorithms (BET, 3dSkullStrip, BSE, optiBET, ROBEX, BEaST).

Results

For the T1-w MRI, our ANN shows significantly higher performance (p<0.001 for 59/60 tests, corrected for multiple comparisons using the family-wise error rate) in both DICE coefficient (97.6) and Hausdorff distance (2.8) compared to all standard methods (DICE coefficient: 94.0–96.0, Hausdorff distance: 4.2–26.0). Moreover, 89.9% of the tests revealed a large and the remaining a medium effect size. Also high performance was shown for brain extraction on T2-w, FLAIR- and cT1-w MRI in both DICE coefficient (95.5–96.9) and Hausdorff distance (3.2–5.0). The improvement yielded with our ANN algorithm as compared to all competing algorithms within the different datasets ranged from +1.65 to +2.66 for DICE and -1.13 to -3.30 for the Hausdorff distance and was most pronounced in the EORTC-26101 dataset (with an increase in the DICE coefficient by +2.66 and decrease in Hausdorff distance by -3.30).

Discussion

Here we present an ANN that enables rapid, automated and robust brain extraction irrespective of pathology, hardware, sequence type or acquisition parameters. We demonstrate generalizability of our ANN within the EORTC-26101 dataset across 37 institutions including all major MRI vendors with a broad variety of scanner types and field strengths as well as within three independent public datasets. Our ANN algorithm yields state-of-the-art performance and outperforms six standard brain extraction algorithms in each of the four datasets. This finding reflects the limitations of existing standard brain extraction algorithms which are not optimized for processing heterogeneous imaging data with abnormal pathologies or varying acquisition parameters and consequently may introduce errors in downstream analysis of neuroimaging data. We addressed this within our study by training (and independent testing) the ANN with data from a large multicentric clinical trial in neuro-oncology which allowed to design a robust and broadly applicable brain extraction algorithm that enables high-throughput processing of neuroimaging data. Moreover, the proposed ANN is able to perform brain extraction on any given anatomical MRI sequence without prior knowledge of the sequence type. From a practical point of view this is of particular importance since image protocols (and the types of sequences acquired) may vary substantially. Standard brain extraction algorithms are essentially optimized to process T1-w MRI sequences whereas they fall short during processing of other types of MRI sequences (e.g. T2-w, FLAIR or cT1-w images). We addressed this shortcoming and demonstrate that our ANN algorithm performs well on cT1-w, FLAIR or T2-w MRI and closely replicates the performance observed for brain extraction on T1-w sequences.

Conclusion

In conclusion, we developed and rigorously validated an ANN that enables rapid, automated and robust brain extraction irrespective of pathology, sequence type, hardware or acquisition parameters and lays the groundwork for automated, high-throughput processing of neuroimaging data.

Acknowledgements

No acknowledgement found.

References

No reference found.