AIMS-TBI - Automated Identification of Moderate-Severe Traumatic Brain Injury Lesions

Moderate to Severe Traumatic Brain Injury (msTBI) is caused by external forces (eg: traffic accidents, falls, sports) causing the brain to move rapidly within the skull, resulting in complex pathophysiological changes. Primary injuries arising from the initial forces (e.g., haematoma, hemorrhages, and contusions)(Mckee & Daneshvar,  2015),  subsequently induce a cascade of secondary injuries (e.g.,gliosis, encephalomalacia, (Maas et al., 2008)) that can include life threatening disorders such as raised intracranial pressure,requiring acute surgical intervention (Bullock et al., 2006). Each of these primary, secondary and surgery related processes can cause structural deformation in the brain. Each patient with msTBI has a unique accumulation of these structural changes, contributing to extremely heterogeneous lesions, considered a hallmark of msTBI  (Covington & Duff, 2021). These lesions differ from other common brain pathologies (stroke, multiple sclerosis (MS), brain tumor) as they can be both focal or diffuse, varying in size, number and laterality, extending through multiple tissue types (e.g., grey matter (GM), white matter (WM), cerebrospinal fluid (CSF)), and can also occur in homologous regions of both hemispheres. Lesions such as these can complicate image registration, normalization, and are known to introduce both local and global errors in brain parcellation (Diamond et al., 2020; King et al., 2020). While several tools exist to compensate for lesions in neuroimaging pre-processing (i.e., the high definition brain extraction tool (HD_Bet) (Isensee et al., 2019), virtual brain grafting (VBG)(Radwan et al., 2021)), most require time consuming manual creation of lesion masks and subsequent manual quality assessment. Furthermore, in our experience, automated methods that have been developed for lesions of different etiologies (e.g. stroke, tumors [Sanjuán et al., 2013]) often underperform in TBI cases.

In the absence of appropriate processing tools, current lesion compensation techniques in msTBI include ignoring the lesions (resulting in unreliable findings), excluding msTBI patients with large lesions (limiting the generalizability), or manual segmentation of lesions prior to analysis (time consuming). Additionally, manual segmentation is often only feasible in smaller, single-site studies which lack the statistical power to perform subgroup analyses. These restrictive approaches limit the ability to investigate how factors such as type of injury (axonal injury, focal lesions, and diffuse microlesions), severity (mild, moderate, severe), and presence of comorbid injuries or complications (especially those that affect pulmonary and cardiovascular function, see [Crawford et al., 2019] and post-traumatic seizure [Bennett et al., 2017; Liesemer et  l., 2011]) impact patient’s functional outcomes. To capture this information, msTBI researchers need access to an accurate, automatic lesion segmentation algorithm trained on large multicohort magnetic resonance imaging (MRI) datasets, where images are aggregated across sites. Whilst a handful of TBI specific algorithms exist, they require either multiple image types (Kamnitsas et al., 2017) (T1-weighted, T2-weighted, fluid attenuated inversion recovery image [FLAIR], gradient echo [GE] & proton density [PD]) or can run on only computed tomography [CT] images (Jain et al., 2019). The necessity for multiple image types limits the applicability of such tools to large-scale consortia aggregating common MRI scans across sites.

The AIMS-TBI challenge leverages data shared within the Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) Consortium (Thompson et al., 2022) Traumatic Brain Injury working group, specifically, the Pediatric msTBI and Adult msTBI subgroups. The AIMS-TBI challenge will focus on identifying lesions in T1-weighted (T1w) MRI data only as it is both the most common MRI scan across our ENIGMA TBI consortium and exhibits less parameter variation (e.g. 1 mm3 voxel size was relatively common in our previous published work, Dennis et al., 2023; Keleher et al., 2022) compared to other MRI modalities (e.g., diffusion MRI) which therefore results in more comparable data collected across sites. The goal of AIMS-TBI is to generate algorithms that can accurately detect and segment 3D lesions, defined as structural damage visible on T1-w MRI due to TBI that may include contusions, hemorrhage, hematoma, encephalomalacia, gliosis, white matter lesions, and surgical drainage tracts.

The inaugural AIMS-TBI Challenge was held at the 2024 MICCAI conference. The final AIMS-TBI 2024 dataset comprised a total of 764 images, including 388 images for training, 101 for validation, and 275 for testing. In the validation phase of the challenge, 12 submissions were uploaded, with 5 teams submitting final models in the test stage. The best performing model achieved an average Dice score of 0.61. This result represented a successful first challenge, yet also highlighted substantial room for improvement. This years’ challenge boasts a larger training dataset to enable more accurate lesion segmentation. Precise lesion segmentation is required for advanced image processing and analyses (such as parcellation, functional connectivity analyses, connectomics, and fixel-based analysis) which hold potential for improving prediction of patient outcomes.

References

• Bennett, K. S., DeWitt, P. E., Harlaar, N., & Bennett, T. D. (2017). Seizures in Children With Severe Traumatic Brain Injury. Pediatric Critical Care Medicine: A Journal of the Society of Critical Care Medicine and the World Federation of Pediatric Intensive and Critical Care Societies, 18(1), 54–63.
• Bullock, M. R., Chesnut, R., Ghajar, J., Gordon, D., Hartl, R., Newell, D. W., Servadei, F., Walters, B. C., & Wilberger, J. E. (2006). Guidelines for the Surgical Management of Traumatic Brain Injury Author Group: Acknowledgments. Neurosurgery, 58(3), S2.
• Covington, N. V., & Duff, M. C. (2021). Heterogeneity Is a Hallmark of Traumatic Brain Injury, Not a Limitation: A New Perspective on Study Design in Rehabilitation Research. American Journal of Speech-Language Pathology / American Speech-Language-Hearing Association, 30(2S), 974–985.
• Crawford, A. M., Yang, S., Hu, P., Li, Y., Lozanova, P., Scalea, T. M., & Stein, D. M. (2019). Concomitant chest trauma and traumatic brain injury, biomarkers correlate with worse outcomes. The Journal of Trauma and Acute Care Surgery, 87(1S Suppl 1), S146–S151.
• Dennis, E. L., Baron, D., BartnikOlson, B., Caeyenberghs, K., Esopenko, C., Hillary, F. G., Kenney, K., Koerte, I. K., Lin, A. P., Mayer, A. R., Mondello, S., Olsen, A., Thompson, P. M., Tate, D. F., & Wilde, E. A. (2020). ENIGMA brain injury: Framework, challenges, and opportunities. In Human Brain Mapping. https://doi.org/10.1002/hbm.25046
• Dennis, E. L., Vervoordt, S., Adamson, M. M., Amiri, H., Bigler, E. D., Caeyenberghs, K., Cole, J. H., Dams-O’Connor, K., Evelyn M Deutscher, Dobryakova, E., Genova, H. M., Grafman, J. H., Håberg, A. K., Hollstrøm, T., Irimia, A., Koliatsos, V. E., Lindsey, H. M., Livny, A., Menon, D. K., … Hillary, F. G. (2023). Accelerated Aging after Traumatic Brain Injury: an ENIGMA Multi-Cohort Mega-Analysis. In bioRxiv (p. 2023.10.16.562638). https://doi.org/10.1101/2023.10.16.562638
• Diamond, B. R., Donald, C. L. M., Frau-Pascual, A., Snider, S. B., Fischl, B., Dams-O’Connor, K., & Edlow, B. L. (2020). Optimizing the accuracy of cortical volumetric analysis in traumatic brain injury. MethodsX, 7, 100994. Henschel, L., Conjeti, S., Estrada, S., Diers, K., Fischl, B., & Reuter, M. (2020). FastSurfer - A fast and accurate deep learning based neuroimaging pipeline. NeuroImage, 219, 117012.
• Isensee, F., Schell, M., Pflueger, I., Brugnara, G., Bonekamp, D., Neuberger, U., Wick, A., Schlemmer, H.-P., Heiland,S., Wick, W., Bendszus, M., Maier-Hein, K. H., & Kickingereder, P. (2019). Automated brain extraction of multisequence MRI using artificial neural networks. Human Brain Mapping, 40(17), 4952–4964.
• Jain, S., Vyvere, T. V., Terzopoulos, V., Sima, D. M., Roura, E., Maas, A., Wilms, G., & Verheyden, J. (2019). Automatic Quantification of Computed Tomography Features in Acute Traumatic Brain Injury. Journal of Neurotrauma, 36(11), 1794–1803.
• Kamnitsas, K., Ledig, C., Newcombe, V. F. J., Simpson, J. P., Kane, A. D., Menon, D. K., Rueckert, D., & Glocker, B. (2017). Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis, 36, 61–78.
• Keleher, F., Lindsey, H. M., Kerestes, R., Amiri, H., Asarnow, R. F., Babikian, T., Bartnik-Olson, B., Bigler, E. D., Caeyenberghs, K., Esopenko, C., Ewing-Cobbs, L., Giza, C. C., Goodrich-Hunsaker, N. J., Hodges, C. B., Hoskinson, K. R., Irimia, A., Königs, M., Max, J. E., Newsome, M. R., … Dennis, E. L. (2022). Multimodal Analysis of secondary
  cerebellar alterations after pediatric traumatic brain injury. In bioRxiv. https://doi.org/10.1101/2022.12.24.22283926
• King, D. J., Novak, J., Shephard, A. J., Beare, R., Anderson, V. A., & Wood, A. G. (2020). Lesion Induced Error on Automated Measures of Brain Volume: Data From a Pediatric Traumatic Brain Injury Cohort. Frontiers in Neuroscience, 14, 491478.
• Liesemer, K., Bratton, S. L., Zebrack, C. M., Brockmeyer, D., & Statler, K. D. (2011). Early post-traumatic seizures in moderate to severe pediatric traumatic brain injury: rates, risk factors, and clinical features. Journal of Neurotrauma, 28(5), 755–762.
• Maas, A. I. R., Stocchetti, N., & Bullock, R. (2008). Moderate and severe traumatic brain injury in adults. Lancet Neurology, 7(8), 728–741.
• Mckee, A. C., & Daneshvar, D. H. (2015). The neuropathology of traumatic brain injury. Handbook of Clinical Neurology, 127, 45–66.
  Radwan, A. M., Emsell, L., Blommaert, J., Zhylka, A., Kovacs, S., Theys, T., Sollmann, N., Dupont, P., & Sunaert, S. (2021). Virtual brain grafting: Enabling whole brain parcellation in the presence of large lesions. NeuroImage, 229, 117731.
• Thompson, P. M., Jahanshad, N., Schmaal, L., Turner, J. A., Winkler, A. M., Thomopoulos, S. I., Egan, G. F., &  Kochunov, P. (2022). The Enhancing NeuroImaging Genetics through Meta-Analysis Consortium: 10 Years of Global Collaborations in Human Brain Mapping. Human Brain Mapping, 43(1), 15–22.