Unsupervised Semi-automated MRI Segmentation Detects Cortical Lesion Expansion in Chronic Traumatic Brain Injury

Holly J. Freeman^1*Alexander S. Atalay¹Jian Li (Andrew)^1,2Evie Sobczak¹Natalie Gilmore¹Samuel Snider³Brian Healy^4,5Holly Carrington⁶Enna Selmanovic^7,8Ariel Pruyser⁷Lisa Bura⁷David Sheppard⁹David Hunt¹⁰Alan C Seifert¹¹Yelena Bodien^1,12Jeanne Hoffman⁹Christine L Mac Donald¹⁰Kristen Dams-O'Connor^13,7Brian L Edlow^14,15

• ¹Center for Neurotechnology and Neurorecovery, Department of Neurology, Massachusetts General Hospital, Harvard Medical School, Boston, United States
• ²Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, United States
• ³Department of Neurology, Brigham and Women’s Hospital and Harvard Medical School, Boston, United States
• ⁴Massachusetts General Hospital Biostatistics Center, Massachusetts General Hospital, Boston, United States
• ⁵Brigham and Women's Hospital Ann Romney Center for Neurologic Diseases, Boston, United States
• ⁶Department of Rehabilitation and Human Performance,, Icahn School of Medicine at Mount Sinai, New York, United States
• ⁷Department of Rehabilitation and Human Performance, Icahn School of Medicine at Mount Sinai, New York, United States
• ⁸Icahn School of Medicine at Mount Sinai Friedman Brain Institute, New York, United States
• ⁹Department of Rehabilitation Medicine, University of Washington School of Medicine, Seattle, United States
• ¹⁰Department of Neurological Surgery, University of Washington, Seattle, United States
• ¹¹Department of Radiology, Icahn School of Medicine at Mount Sinai BioMedical Engineering and Imaging Institute, New York, United States
• ¹²Department of Physical Medicine and Rehabilitation, Spaulding Rehabilitation Hospital Corporation, Boston, United States
• ¹³Icahn School of Medicine at Mount Sinai Department of Neurology, New York, United States
• ¹⁴Center for Neurotechnology and Neurorecovery, Department of Neurology, Massachusetts General Hospital and Harvard Medical School, Boston, United States
• ¹⁵Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, United States

Traumatic brain injury (TBI) is a risk factor for neurodegeneration and cognitive decline, yet the underlying pathophysiologic mechanisms are incompletely understood. This gap in knowledge is in part related to a lack of reliable and efficient methods for measuring cortical lesions in neuroimaging studies. The objective of this study was to develop a semi-automated lesion detection tool and apply it to an investigation of longitudinal changes in brain structure among individuals with chronic TBI. We identified 24 individuals with chronic moderate-to-severe TBI enrolled in the Late Effects of TBI (LETBI) study who had cortical lesions detected by T1-weighted MRI and underwent two MRI scans at least two years apart. Initial MRI scans were performed more than one year post-injury, and follow-up scans were performed 3.1 (IQR=1.7) years later. We leveraged FreeSurfer parcellations of T1-weighted MRI volumes and a recently developed super-resolution technique, SynthSR, to automate the identification of cortical lesions in this longitudinal dataset. Trained raters received the data in a randomized order and manually edited the automated lesion segmentations, yielding a final semi-automated lesion mask for each scan at each time point. Inter-rater variability was assessed in an independent cohort of 10 additional LETBI subjects with cortical lesions. The semi-automated lesion segmentations showed a high level of accuracy compared to "ground truth" lesion segmentations performed via manual segmentation by a separate blinded rater. In a longitudinal analysis of the semi-automated segmentations, lesion volume increased between the two time points with a median volume change of 4.91 (IQR=12.95) mL (p<0.0001). Lesion volume significantly expanded in 37 of 61 measured lesions (60.7%), as defined by a longitudinal volume increase that exceeded inter-rater variability. Longitudinal analyses showed similar changes in lesion volume using the ground-truth lesion segmentations. Inter-scan duration was not associated with the magnitude of lesion growth. While the proposed tool requires further refinement and validation, we show that reliable and efficient semi-automated lesion segmentation is feasible in studies of chronic TBI, creating opportunities to elucidate mechanisms of post-traumatic neurodegeneration.