INTRODUCTION

In neurosurgical preoperative planning of resectionof large tumors it is important to locate the paths of critical cerebral nerve fiber bundles (e.g. corticospinal tract). This information can then be incorporated into the surgical procedure. Using MR diffusion-weighted imaging nerve fiber bundles can be virtually reconstructed using probabilistic or deterministic tracking procedures. The subsequent proper selection of the bundles of interest is often performed manually¹ (ROI selection) or utilizes atlas information². The first approach requires expert knowledge and is susceptible to user errors, whereas the second approach is often not well suited or has not been tested in the presence of large brain tumors including accompanying brain edema and tissue shifts. Recently, a novel framework was introduced for fast and accurate white matter tract segmentation³ (TractSeg) using a deep neural network. The method does not rely on virtual fiber bundle reconstructions, but performs the segmentation by combining several 2D segmentations based on reconstructed ODF peak information. In this study, we investigated the feasibility of this novel fully automatic segmentation approach to identify white matter fiber bundles in the presence of large brain masses.

METHODS

This study included 12 data sets of patients (male= 9, age range: 26-72) with different types of glioma (n=8), brain metastases (n=2) or arteriovenous malformations (n=1). The pre-surgical MRI scans included at least one high resolution (0.7 mm iso) anatomical whole-head scan (MP-RAGE or T2-SPACE) and a diffusion weighted imaging scan (1.5 mm iso, 92 slices, SMS factor 2, 4xb0, 30x different diffusion directions @ b=1000 s/mm^2). The DWI scan was performed twice with reversed phase encoding (AP+PA). All measurements were acquired on a Siemens Skyra 3T MRI using a 20-ch head coil. The reconstruction pipeline is illustrated in Fig. 1 and included software components from FSL⁴, Mrtrix3, TractSeg as well as some functionality implemented in Matlab (The Mathworks, Nattik, USA). The data were denoised, unwarped, corrected for eddy current based deformations and co-registered to the anatomy scan. The pre-processed DWI data was fed into the TractSeg pipeline. The tract probability maps were generated in MNI space and then transformed back to the anatomical space to be overlayed onto the anatomical image data. Tract probability maps were superimposed as heat maps onto the gray-scale anatomy scan and stored with image position information in DICOM format and send to the hospital PACS for review. An experienced neuro-radiologist inspected the overlays in coronal, axial and sagittal orientation and evaluated the plausibility of the reconstructed fiber bundle (left+right corticospinal tract).

RESULTS

The corticospinal tracts could be reconstructed in both brain hemispheres for all 12 cases, although in almost all cases at least one hemisphere was substantially deformed, misplaced or infiltrated by tumor tissue. The latter case was usually accompanied with reduced probabilities in the overlay maps. The neuro-radiologist assessed all superimposed data sets as being plausible reconstructions of the corticospinal tracts connecting the brainstem with the pre-central gyrus. Examples of the overlay maps are shown in Fig. 2-5. For two patients, the anatomical scan suffered from motion induced artifacts; nevertheless, the quality was still sufficient to assess the plausibility of the reconstructed tracts.

DISCUSSION

Applying the proposed complex, but fully automatic approach in patients with large brain tumors or other brain tissue shifting lesions for identifying individual important fiber bundles in pre-surgical planning, revealed promising results. The pre-trained deep neural network, which had been trained on healthy volunteer data acquired with a much more advanced diffusion imaging protocol3, was successfully applied to the patient data without any modifications. The results demonstrate the general applicability of the framework, especially for handling of data from different acquisition protocols. Since there is no ground truth available to identify the nerve fiber bundles in vivo, a final validation of this approach will be difficult. Its robustness against brain tissue shifts might be investigated systematically by using digital phantomsor with intraoperative navigated stimulation.

CONCLUSION

Although a ground truth method to validate the presented results in-vivo is missing, the proposed results were rated as plausible reconstructions by an experienced neuro-radiologist. Thus the proposed processing pipeline could become a valuable tool for pre-surgical planning. It might also find application in radiation therapy planning for providing additional information as well as in treatment monitoring. Future research should address – besides validation – potential influencing factors affecting the reliability of the reconstructed results, such as spatial resolution or diffusion direction information.

Acknowledgements

No acknowledgement found.