In this prospective study of all TBI severities with early MRI, we investigated the location, number, and volume of TAI lesions as potential predictors for outcome, after adjusting for established outcome predictors. In severe TBI, the presence of bilateral TAI in the brainstem or thalami was a strong outcome predictor, especially when located in pons. Interestingly, in mild-moderate TBI, the total volume of contusions on MRI was more important for outcome than TAI volume. Based on our results, we propose the Trondheim TAI-MRI grading (Fig. 4a, b) that can be applied by visual evaluation of early MRI. In all TBI severities, however, the best model fit was found when quantitative FLAIR models replaced the TAI-MRI grading.

We found that bilateral TAI in pons most strongly predicted worse outcomes at 6 months in severe TBI with ORs among the highest across the studied locations. In a retrospective MRI study of 255 critically ill TBI patients, the presence of bilateral TAI in pons was also proposed to represent the worst grade [12]. In a study from 2002, not specifically studying TAI, any bilateral MRI lesions in upper pons were the strongest predictor for mortality [29]. We therefore propose bilateral TAI in pons as the worst grade, Trondheim TAI-MRI grade 5.

The presence of bilateral TAI in mesencephalon or thalami was a strong outcome predictor in severe TBI. The thalamus consists mainly of grey matter nuclei but is surrounded by layers of WM and separated by a Y-shaped layer of WM, the internal medullary lamina [30], which may explain why TAI can be found in the thalamus. The thalamus is an important relay centre with reciprocal connections to nearly all parts of the brain, with the intralaminar nuclei embedded in the internal medullary lamina, particularly important for consciousness [30]. This can explain why bilateral TAI in the thalami is so important for the outcome. Also, we have previously found that patients with bilateral TAI in the thalami had lower GOSE scores than those with unilateral TAI in the thalamus [11], and bilateral TAI in the thalami was far more indicative of a low admission GCS than any other MRI finding [13]. In a DTI study, we found lower fractional anisotropy values in the thalamus in all standard TAI grades [31]. Also, two recent reviews on MRI in TBI concluded that any bilateral lesions in the brainstem or thalami increased the risk for poor outcomes [32, 33]. Finally, patients with bilateral TAI in the thalami were also associated with poor outcomes in the Stockholm MRI grading [12]. We propose that bilateral TAI in mesencephalon or thalami should be classified as Trondheim TAI-MRI grade 4.

Further, in all TBI analysed together, we found that unilateral TAI in the thalamus or brainstem and bilateral TAI lesions in the basal ganglia significantly predicted worse outcomes. It was expected that unilateral TAI in the brainstem or thalamus was not as important for the outcome as bilateral injuries, but we found it somewhat surprising that bilateral TAI in basal ganglia was not so closely associated with poor outcomes. However, the basal ganglia is primarily involved in motor control, while the brain stem and thalami are more important for vital functions and consciousness [11, 30]. We recently found that the presence of unilateral TAI in the brainstem was significantly associated with GCS score [13]. In moderate TBI with a GCS score of 9–12, the presence of any TAI in the brainstem, thalamus, or basal ganglia was not significantly associated with the outcome. However, the estimated OR (2.9) was similar to the one for severe TBI (OR 3.1), and the lower degree of evidence of an effect on outcome might be due to the lower frequency of TAI. Importantly, no patients with mTBI had TAI in the brainstem, thalamus, or basal ganglia. In a retrospective MRI study of 178 patients with severe TBI and TAI, multivariable ordinal regressions with adjustment for IMPACT variables also demonstrated the importance of any TAI in thalamus/basal ganglia for outcome at 12 months, in addition to TAI in the corpus callosum and brainstem [34]. We propose that the presence of unilateral TAI in the thalamus or brainstem or unilateral/bilateral TAI in basal ganglia should be classified as Trondheim TAI-MRI grade 3.

The presence of TAI in the corpus callosum was significantly associated with outcomes in severe TBI. However, many of these patients also had TAI in the brainstem, thalamus, or basal ganglia and it is difficult to deduct the contribution to outcome prediction. We did not find any evidence that TAI in the splenium was a stronger predictor of worse outcomes than TAI in genu/truncus, in contrast to the observed association with GCS score [13]. For clinical purposes, we therefore suggest that TAI in the corpus callosum is not further subdivided and is classified as grade 2. We also suggest that TAI in hemispheres or cerebellum still should be classified as grade 1, since there was little evidence in our data to recommend changing the current practice. Patients with mTBI almost exclusively only had TAI in the hemispheres; and in a larger sample, it is reasonable to anticipate that such lesions will be associated with outcomes even though we could not demonstrate a statistically significant effect.

In severe TBI, the Trondheim TAI-MRI grading performed better in predicting 6-month outcomes compared to the standard TAI grading, the Stockholm MRI grading [12] as well as our TAI-MRI grading based on GCS score [13]. The Stockholm MRI grading has a higher number of sublocations included in their grades 2 and 3, while the Trondheim TAI-MRI grading is more similar to the standard grading used today and thereby easier to learn and implement for the radiologist in everyday clinical practice. We also question that the Stockholm MRI grading does not distinguish patients without TAI on MRI from patients with TAI in hemispheres, since both will be allocated to grade 1 in that grading system.

In msTBI, the total volumes of TAI were more important outcome predictors than the total numbers, and volumes on FLAIR were more important than on DWI. Adjusting for the time factor on DWI did not improve model fit. We know from stroke imaging that DWI lesions disappear or attenuate 2–3 weeks after ictus [35], which is also the clinical experience in TBI. Thus, DWI is less useful in a clinical setting since MRI is typically performed later in msTBI than in stroke.

The prognostic model including TAI-FLAIR volumes gave high model fit in msTBI. The importance of TAI-FLAIR volumes in msTBI is supported by other smaller studies [10, 36, 37]. Interestingly, in all TBI severities, we found a better model fit generally when quantitative models replaced the clinical TAI-MRI grading. In moderate (GCS score 9-12) and severe TBI, the TBI-FLAIR volume model (including volume of TAI and contusion) gave the highest model fit, while in GCS score 13–15 the Contusion-FLAIR volume model contributed to the highest model fit. Smaller studies have earlier shown the predictive value of contusions in moderate [10] and mTBI [38]. The finding that FLAIR volumes gave even higher model fit than clinical MRI gradings, is promising for the use of artificial intelligence (AI) technologies. However, also in our models, a large proportion of the variance in the GOSE score remained unexplained. The outcome after TBI is multidimensional and assumed to be influenced not only by injury severity but also by other factors such as contextual factors and psychosocial functioning.

This study has several strengths: First, the prospective data collection and the large number of patients with early MRI. Second, we performed extensive structured template-based MRI readings and manual lesion segmentations on three different MRI sequences. Manual segmentation is regarded as the gold standard, automatic algorithms are promising but still not available for independent use [39]. Third, the MRI readings and segmentations were all performed blinded and quality-checked in inter-rater-analyses with good inter-rater-agreement [13].

One limitation is the selection bias that always will be present in early MRI studies of TBI, and we have earlier acknowledged reasons for this, such as age and injury severity [11, 14]. Even though the total sample is large, the lower number of patients together with the lower prevalence of MRI findings result in lower power in moderate and particularly mTBI. Another limitation is the heterogeneity of the MRI scanners with most patients examined with 1.5 T scanners in msTBI when preferably the whole cohort should have been imaged on 3 T. However, in a clinical setting, both 1.5 T and 3 T scanners will be used many years ahead and it is beneficial with a grading that can be used independently of field strength. Many of the msTBI patients in this cohort were examined with T2*GRE instead of SWI, which may have led to an underestimation of TAI. Thus, we recommend that the Trondheim TAI-MRI grading and the quantitative models will be externally validated in upcoming larger multicentre datasets with 3 T and SWI.

In conclusion, we propose the Trondheim TAI-MRI grading, with bilateral TAI in mesencephalon or thalami and bilateral TAI in pons as the worst grades 4 and 5, respectively. The Trondheim TAI-MRI grading most reliably estimated outcome in severe TBI, larger sample sizes will be necessary to clarify the importance in mild-moderate TBI. Interestingly, TAI seemed to be less important for outcome prediction in mild-moderate TBI where the volume of contusions on MRI predicted outcome better. The quantitative models comprising FLAIR lesion volumes, had the highest model fits in all TBI severities. In the future, the continuous improvements of AI will likely enable the use of quantitative models in the clinic. A more optimal prognostic classification of brain injury on early MRI will be important to help decision-making, informing patients and families, and stratifying patients for optimal follow-up.