Quantitative imaging biomarkers of damage to critical memory regions associated with post-radiotherapy memory performance in brain tumor patients

Abstract

Purpose:

We used quantitative MRI to prospectively analyze the association between microstructural damage to memory-associated structures within the medial temporal lobe and longitudinal memory performance after brain radiotherapy (RT).

Methods and Materials:

Primary brain tumor patients receiving fractionated brain RT were enrolled on a prospective trial (n=27). Patients underwent high resolution volumetric brain MRI, diffusion-weighted imaging, and neurocognitive testing prior to and 3, 6, and 12 months post-RT. Medial temporal lobe regions (hippocampus; entorhinal, parahippocampal, and temporal pole white matter [WM]) were auto-segmented, quantifying volume and diffusion biomarkers of WM integrity (mean diffusivity [MD]; fractional anisotropy [FA]). Reliable change indices (RCI) measured changes in verbal (Hopkins Verbal Learning Test-Revised [HVLT-R]) and visuospatial (Brief Visuospatial Memory Test-Revised [BVMT-R]) memory. Linear mixed-effects models assessed longitudinal associations between imaging parameters and memory.

Results:

Visuospatial memory significantly declined at 6 months post-RT (mean RCI - 1.3, P=.012). Concurrent chemotherapy and seizures trended toward a significant association with greater decline in visuospatial memory (P=.053, P=.054, respectively).

Higher mean dose to the left temporal pole WM was significantly associated with decreased FA (r=−.667, P=.002). Over all time points, smaller right hippocampal volume (P=.021), lower right entorhinal FA (P=.023), greater right entorhinal MD (P=.047), and greater temporal pole MD (BVMT-R Total Recall, P=.003; BVMT-R Delayed Recall, P=.042) were associated with worse visuospatial memory. The interaction between right entorhinal MD (BVMT-R Total Recall, P=.021; BVMT-R Delayed Recall, P=.004), and temporal pole FA (BVMT-R Delayed Recall, P=.024) significantly predicted visuospatial memory performance.

Conclusions:

Brain tumor patients exhibited visuospatial memory decline post-RT. Microstructural damage to critical memory regions, including the hippocampus and medial temporal lobe WM, were associated with post-RT memory decline. The integrity of medial temporal lobe structures is critical to memory performance post-RT, representing possible avoidance targets for memory preservation.

INTRODUCTION

Radiation therapy (RT) is a mainstay in the treatment of benign and malignant primary brain tumors. However, RT can result in post-treatment neurocognitive decline¹, most frequently reported in verbal and visuospatial memory (i.e., difficulty encoding, retaining, and retrieving visual information). Neurocognitive decline has been shown to be an independent predictor of survival for patients with brain tumors, and the late delayed (6 months and greater) effects of RT are typically irreversible and progressive¹. Thus, neurocognitive outcomes have become a critical endpoint in brain RT clinical trials².

Radiation-induced injury to the brain is mediated by destruction of actively dividing progenitor cells (e.g. hippocampus), damage to white matter (WM) tracts, vascular injury, and neuroinflammation¹. Axonal degradation and demyelination of WM has been noted on histopathologic studies after radiation exposure^3,4, and diffusion tensor imaging (DTI) biomarkers are associated with these changes. Indeed, previous DTI studies have shown regional dose-dependent WM damage^5–8 after RT: specifically, a decrease in Fractional Anisotropy (FA) and increase in Mean Diffusivity (MD) indicating loss of white matter integrity. Other studies have demonstrated associations between WM damage to certain structures (i.e., parahippocampal cingulum) and neurocognitive decline^9–11. In addition to WM, gray matter structures, such as the hippocampus, exhibit dose-dependent atrophy¹² after RT and likely play a role in radiation-induced neurocognitive decline¹³.

In this prospective study of primary brain tumor patients receiving fractionated brain RT, we analyzed discrete gray and WM regions involved in memory using high-resolution structural and diffusion-weighted imaging. Specifically, we investigated the association between structural and microstructural damage to medial temporal lobe regions and post-RT decline in both verbal and visuospatial learning and memory. This study is the first among those investigating adult patients with brain tumors to include visuospatial memory as an endpoint. The ultimate objective of this work is to better understand the neuroanatomic regions involved in post-RT memory decline and identify potential targets for memory-sparing brain RT.

METHODS AND MATERIALS

Standard Protocol Approvals and Patient Consents

This study was approved by our institutional review board. All participants provided written informed consent prior to participation.

Study Design and Participants

Adults with primary brain tumors who were eligible for fractionated partial-brain RT with protons or photons (1.8-2.0 Gy per fraction, 50.4-60 Gy total dose) were enrolled in the parent prospective, observational study from 2014-2016. Eligibility criteria included age >18, Karnofsky performance status >70, ability to answer questions and follow commands in English at the time of consultation and treatment, and estimated life expectancy >1 year. Patients who received prior brain RT were excluded. Patients were studied at four time points: baseline (pre-RT), 3 months, 6 months, and 12 months post-RT. At each time point, high-resolution 3D volumetric brain MRI and diffusion-weighted imaging (DWI) were obtained (per clinical standard-of-care at our institution), and a battery of neurocognitive tests was administered.

Memory Assessment

All participants underwent a 2-hour neuropsychological assessment at each time point. The battery included well-validated measures of verbal and visuospatial memory, executive functioning, and attention/processing speed, the domains that demonstrate greatest impairment in patients after brain RT¹⁴. The current analysis focused on a subset of the tests assessing verbal (Hopkins Verbal Learning Test-Revised [HVLT-R])¹⁵ and visuospatial memory (Brief Visuospatial Memory Test-Revised [BVMT-R])¹⁶. These tests have alternate forms, which are ideal for repeat testing as different, but psychometrically equivalent, forms were used at each of the four time points to avoid patient “learning” of the tests. We analyzed structure-function associations that have been validated in previous studies (see eTable 1)^17–19.

Imaging Acquisition

High-resolution volumetric and diffusion-weighted MRI scans for all patients at each time point were acquired on a 3.0T 750 GE system (GE Healthcare, Milwaukee, Wisconsin) equipped with an 8-channel head coil. The imaging protocol included a 3D volumetric T1-weighted inversion recovery spoiled gradient echo sequence (echo time [TE]/repetition time [TR] = 2.8/6.5 ms; inversion time [TI] = 450 ms; flip angle = 8 degrees; field of view [FOV] = 24 cm) and a 3D FLAIR sequence (TE/TR = 125/6000 ms, TI = 1868 ms, FOV = 24 cm, matrix = 256×256, slice thickness = 1 mm). DWI was acquired with a single-shot pulsed-field gradient spin EPI sequence (TE/TR = 96 ms/17 s; FOV = 24 cm, matrix = 128 × 128 × 48; 1.87 × 1.875 in-plane resolution; slice thickness = 2.5 mm; 48 slices) with b = 0, 500, 1500, and 4000 s/mm2, with 1, 6, 6, and 15 unique gradient directions for each b-value respectively, and one average for each non-zero b-value. For use in nonlinear B₀ distortion correction, two additional b=0 volumes were acquired with either forward or reverse phase-encode polarity.

Image Processing

The imaging data were preprocessed using in-house algorithms developed in MATLAB. Anatomical images were corrected for distortions due to gradient nonlinearities²⁰. Diffusion scans were corrected for spatial distortions associated with gradient nonlinearities, susceptibility, and eddy currents^21,22. FreeSurfer, 5.3.0 was used to parcellate volumetric MRI into 34 cortical gyral-based ROIs²³. Diffusion tensor imaging (DTI) maps of fractional anisotropy (FA) and mean diffusivity (MD) were derived by fitting the DWI data from b-values of 0, 500, and 1500 s/mm² to a tensor. In each voxel, the diffusion process is approximated by an ellipsoid defined by three perpendicular axes or eigenvectors. MD is a rotationally invariant measure of the average mobility of water molecules, calculated as an average of the three eigenvalues, and expressed as mm²/sec. FA ranges from 0 to 1 as an expression of the degree of directional bias of diffusion.

Next, these DWI-derived maps were co-registered to high resolution volumetric MRI, and FA and MD values within the superficial white matter were calculated by sampling 5 mm below the WM surface normal at each vertex and then averaged within each ROI. Selected medial temporal lobe ROIs included the hippocampus and entorhinal, parahippocampal, and temporal pole WM (eTable 1; eFigure 1). These ROIs were preselected based on prior research implicating their role in the associated neurocognitive domains^{17–19,24–26}. To avoid measuring tumor- or edema-related effects, a censoring mask including tumor, tumor bed, surgical cavity, surgical scars, and edema (T2 FLAIR hyperintensity) was manually drawn slice-by-slice on each image, and verified by two imaging experts. Voxels in the censoring mask were excluded from the final ROI to avoid confounding by tumor and edema-related effects⁶.

End Points

Change in each imaging parameter and memory measure were evaluated in two ways to capture the following: 1) subacute effects (i.e. change from baseline to 6 months post-RT), and 2) longitudinal evaluation of subacute and late-delayed effects, encompassing all time points (baseline, 3, 6, and 12 months post-RT). Reliable change indices (RCIs) were used to calculate neuropsychological change from baseline to 6 months post-RT, accounting for practice effects²⁷. To evaluate time-dependent longitudinal performance, raw test scores and imaging parameters (FA and MD for WM, volume for gray matter) were analyzed. Raw neurocognitive test scores were used as opposed to age/education adjusted T scores so that we could independently investigate any associations between age, education, and outcome.

Statistical Analysis

Statistics were performed in SPSS v24 (IBM Corp). To assess baseline to 6 months post-RT change, each parameter was evaluated by a one-sample t-test (H₀=0). Associations between subject characteristics (i.e. demographics, tumor type, chemotherapy) and 6-month changes in memory outcomes using RCIs were evaluated by Pearson correlations, independent sample t-tests, and one-way ANOVA.

To assess the main effects of time and imaging parameters as predictors of memory performance, random intercept and slope linear mixed effects (LME) models were performed:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_{\text{0}}{\text{+b}}_{0\mathrm{i}})+({\beta }_{\text{1}}{\text{+b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\beta }_2{\mathrm{Imaging}+\mathrm{e}}_{\mathrm{ij}}$$

where b_0i = subject-specific random intercept, b_1i = random slope, and e_ij = subject error. This model was used to account for within-subject correlation between repeated measures, random subject intercepts, and incomplete outcomes (i.e. some patients were missing certain memory tests at certain time points, Supplemental eTable 2a). A random time component was specifically included to account for the change over time of the subjects.

In a separate analysis, the interaction between time and imaging parameters was included to evaluate whether the association between imaging and memory performance changes with time:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_{\text{0}}{\text{+b}}_{0\mathrm{i}})+({\beta }_{\text{1}}{\text{+b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\beta }_2{\mathrm{Imaging}+\beta }_3\mathrm{Imaging}\times {\mathrm{Month}}_{\mathrm{j}}{+\mathrm{e}}_{\mathrm{ij}}$$

Outliers were identified and removed via Mahalanobis distance based on a chi-square distribution (assessed using P<.001)²⁸ Statistical significance was set at α=0.05 for two-tailed tests.

Post-hoc analyses were done to further investigate time trends of neurocognitive performance in addition to evaluating baseline variability of particularly relevant clinical variables, namely age and tumor type. Based on significant baseline variability using independent sample t-tests, further analyses were performed including the significant baseline variables in the linear mixed effects models.

RESULTS

Subacute Effects: Baseline to 6-Months Post-RT

Of the 56 subjects enrolled on the trial, 22 subjects had both pre- and 6 months post-RT imaging and memory assessments and were included in this analysis of subacute (baseline to 6-months post-RT) effects (eFigure 2). All subjects’ raw scores at each time point are shown in eFigure 3.

Patient demographics and treatment characteristics are shown in Table 1. Notably, most patients were Non-Hispanic White. Most patients had a diagnosis of glioma and the minority had proton beam therapy as opposed to standard photon intensity modulated radiation therapy (IMRT). We noted that 14 patients (64%) were on anti-inflammatory corticosteroids and 10 (46%) had seizures at some point during the 6-month follow-up period.

Table 1.

Subject and cancer characteristics (N=22)

Demographic	Patients, No. (%)
Gender
Men	11 (50.0)
Women	11 (50.0)
Race
Non-Hispanic White	18 (81.8)
Black	1 (4.5)
Hispanic	3 (13.6)
Age (median, range)	48 (20, 75)
Education, years (median, range)	14 (10, 20)
Cancer or Treatment Characteristic	Patients, No. (%)
Glioma	14 (63.6)
Laterality
Left	7 (31.8)
Right	10 (45.5)
Bilateral	5 (22.7)
Radiation Therapy
Proton Beam Therapy	7 (31.8)
IMRT	15 (68.2)
Radiotherapy Prescription Dose (median [Gy], range)	54.0 (50.4, 60.0)
Chemotherapy
Concurrent	11 (50.0)
Adjuvant	13 (59.0)
Steroids	12 (54.5)
Seizures	10 (45.5)

Abbreviations: N, number; IMRT, intensity modulated radiation therapy; Gy, Grey

Table 2 shows the mean dose delivered to each ROI. Mean doses to the left and right hippocampi were 13.8 Gy (SD=13.7) and 19.3 Gy (SD=18.0), respectively. Higher mean dose to the left temporal pole WM was significantly associated with decreased FA (r=−.667, P=.002).

Table 2.

Mean dose and imaging parameter by region

Structure	Sample Size* (N)	Mean Dose (mean Gy, SD)	Percent Change in Volume (mean %, SD) / Number Atrophied (%)	Percent Change in MD (mean %, SD)	Percent Change in FA (mean %, SD)
Left hippocampus	18	13.8 (13.7)	−2 (17) / 9 (50)	-	-
Right hippocampus	19	19.3 (18.0)	+4 (28) / 10 (53)	-	-
Left WM
Entorhinal	19	20.4 (20.9)	-	17 (8)	−16 (23)
Parahippocampal	19	22.6 (23.2)	-	−5 (8)	6 (27)
Temporal pole	18	24.8 (20.0)	-	15 (8)	−10 (5)
Right WM
Entorhinal	18	23.8 (18.9)	-	19 (10)	−16 (15)
Parahippocampal	19	24.5 (22.2)	-	−8 (10)	11 (11)
Temporal pole	19	24.8 (20.0)	-	16 (7)	−31 (15)

^*The analytic sample size is less than the total eligible sample size of 22 subjects due to censoring.

Percent change is shown from baseline to 6 months (i.e. a negative value represents a decrease over time). Abbreviations: SD, standard deviation; WM, white matter; Gy, Grey; FA, fractional anisotropy; MD, mean diffusivity.

When grouping by disease (glioma vs no glioma), there were some significant differences in baseline imaging parameters and memory performance. Specifically, patients with gliomas had significantly higher baseline right temporal pole MD values (mean difference +5.25×10⁻⁵, 95% CI [1.76–8.73×10⁻⁵], P=0.005). Patients with gliomas also had worse baseline performances of HVLT Total (mean difference −4.01, 95% CI [6.91, −1.10], P=0.009) and Delayed Recall (mean difference −2.31, 95% CI [−3.95,−0.66], P=0.008, equal variance not assumed).

Mean normalized changes in verbal and visuospatial memory from baseline to 6-months post-RT are summarized in Table 3. BVMT-R Total Recall declined from baseline (mean RCI= −1.34, 95% CI [−2.359, −0.328], P=.012). Older age was correlated with less 6-month decline in HVLT-R Total Recall (r=0.484, P=.026). This association remained significant (P=.021) on post-hoc analysis upon including baseline performance on HVLT-R Total Recall as a covariable. There was a strong trend for patients who had seizures to show a greater decline in BVMT-R Delayed Recall (mean RCI [$\overline{\mathrm{x}}$]= −1.90, SD=2.92) than those without seizures ($\overline{\mathrm{x}}=0.30$, SD=1.98; t[19]=2.06, P=.053). Patients who received concurrent chemotherapy had a trend toward greater decline in visuospatial memory (BVMT-R Total Recall; $\overline{\mathrm{x}}=-2.39$, SD=2.75) than patients not treated with concurrent chemotherapy ($\overline{\mathrm{x}}=-0.40$, SD=1.02;, t[11.25]=2.15, P=.054). Levene’s test indicated unequal variances (F=7.95, P=.011), so degrees of freedom were adjusted from 19 to 11.25.

Table 3.

Change in neurocognitive tests at 6 months post-RT from baseline

Memory Test	Mean RCI (95% CI)	P-value*
Verbal Memory
HVLT-R Total Recall	−0.013 [−0.709, 0.682]	0.968
HVLT-R Delayed Recall	−0.104 [−0.602, 0.394]	0.667
Visuospatial Memory
BVMT-R Total Recall	−1.343 [−2.359, −0.328]	0.012 †
BVMT-R Delayed Recall	−0.645 [−1.835, 0.544]	0.271

Abbreviations: RT, radiotherapy; BVMT-R, Brief Visuospatial Memory Test-Revised; HVLT-R, Hopkins Verbal Learning Test-Revised; RCI, Reliable Change Index

^*P-value represents one sample T test (H₀=0, no change from baseline).

^†Significant at P<.05.

Neuroimaging Biomarkers of Memory Performance

Linear Mixed Effects Analysis of Baseline to 12 Months Post-RT: Main Effects

Of 56 patients enrolled, 27 subjects were eligible for the baseline to 12-month analyses (had at least two time points for both memory assessments and MRI over the 12-month study period). Supplemental eTable 2a provides information on missing data points in the full longitudinal follow-up and eTable 2b details any significant differences in patient characteristics between subjects with and without missing data.

When considering all time points for each patient up to 12 months post-RT, the main effect of left hippocampal volume (β₂ in the linear mixed effects model) was not significantly associated with verbal memory performance (Figure 1). However, smaller right hippocampal volumes were associated with poorer BVMT-R Delayed Recall performance (β₂= 0.00214 points/mm³, P=.021) and trended toward significantly associated with poorer BVMT-R Total Recall (β₂= 0.004 points/mm³, P=.069)

Figure 1.

Scatter plots for hippocampal volume and domain-specific memory performance including all time points for each patient up to 12 months post-RT. The trend line overlays the LOESS fit with the smoothing parameter that minimizes the AICC criterion. Significant associations between imaging parameter and memory test were determined based on the beta coefficient (β₂) derived from the linear mixed effects model with random intercept and slope:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_0{+\mathrm{b}}_{0\mathrm{i}})+({\beta }_1{+\mathrm{b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\beta }_2{\mathrm{Imaging}+\mathrm{e}}_{\mathrm{ij}}$$

Raw memory scores are shown. Hippocampal volumes are shown as mm³.

(A) Smaller left hippocampal volumes were not significantly associated with poorer performance on verbal memory testing (HVLT-R Total Recall β₂= 0.00038, P=.849; HVLT-R Delayed Recall, β₂= 0.00008, P=.935).

(B) Smaller right hippocampal volumes were significantly associated with worse performance on visuospatial memory testing (BVMT-R Total Recall, β₂= 0.004 points/mm³, P=.069; BVMT-R Delayed Recall, β₂= 0.00214 points/mm³, P=.021). Abbreviations: BVMT-R, Brief Visuospatial Memory Test-Revised; HVLT-R, Hopkins Verbal Learning Test-Revised

Neither FA nor MD of left hemispheric WM within the parahippocampal, entorhinal, or temporal pole ROIs were associated with HVLT-R Total or Delayed Recall performance across time points (all P-values > .05). However, lower right entorhinal FA values and higher right entorhinal MD values were significantly associated with worse performance on BVMT-R Total Recall (β2= 49.15 points, P=.023; β₂= −28,385 points/mm²/s, P=.047, respectively) (Figure 2). Higher right temporal pole MD values were significantly associated with poorer visuospatial memory (BVMT-R Total Recall, β₂= −60,800 points/mm²/s, P=.003; BVMT-R Delayed Recall, β₂= −17,762 points/mm²/s, P=.042). The main effect of time was not significantly associated with verbal or nonverbal memory.

Figure 2.

Scatter plots for (A) right entorhinal and (B) right temporal WM and visuospatial memory performance including all time points for each patient up to 12 months post-RT. The trend line overlays the LOESS fit with the smoothing parameter that minimizes the AICC criterion. Significant associations between imaging parameter and memory test were determined based on the beta coefficient (β₂) derived from the linear mixed effects model with random intercept and slope:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_0{+\mathrm{b}}_{0\mathrm{i}})+({\beta }_1{+\mathrm{b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\beta }_2{\mathrm{Imaging}+\mathrm{e}}_{\mathrm{ij}}$$

Raw memory scores are shown. MD is expressed in mm²/s. FA is unitless. Outliers (n=2 and n=4 for A and B, respectively) were removed based on statistically significantly great Mahalanobis distances (P<.001).

(A) Higher right entorhinal MD values were significantly associated with worse BVMT-R Total Recall (β₂= −28,385 points/mm²/s, P=.047). Smaller right entorhinal FA values were significantly associated with worse BVMT-R Total Recall (β₂= 49.15 points,P=.023). (B) Higher right temporal pole MD values were significantly associated with worse nonverbal memory (BVMT-R Total Recall, β₂= −60,800 301 points/mm²/s , P=.003; BVMT-R Delayed Recall, β₂= −17,762 301 points/mm²/s, P=.042).

Abbreviations: BVMT-R, Brief Visuospatial Memory Test-Revised; HVLT-R, Hopkins Verbal Learning Test-Revised

On post-hoc analysis, time was evaluated without accounting for imaging parameters as shown in the following model:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_{\text{0}}{\text{+b}}_{0\mathrm{i}})+({\beta }_{\text{1}}{\text{+b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\mathrm{u}}_{\mathrm{i}}{+\mathrm{e}}_{\mathrm{ij}}$$

HVLT Delayed Recall significantly declined over time (β₁ −0.14, P=0.018) (eFigure 3). When grouping the patients by age (old [60+] vs young), there were no significant differences in either baseline imaging parameters or neurocognitive tests (independent sample t-tests at P<0.05). Of note, despite baseline differences in imaging parameters and memory between patients with and without glioma, glioma was not a significant predictor of neurocognitive performance.

Linear Mixed Effects Analysis of Baseline to 12 Months Post-RT: Time Interaction

The interaction between left hippocampal volume and time was not significantly associated HVLT-R performances (Table 4). Similarly, the interactive effects of right hippocampal volume and time were not significantly associated with either BVMT-R Total Recall or Delayed Recall performance.

Table 4.

Interactions between imaging biomarkers and time

Right-hemispheric Structure	Imaging Biomarker	BVMT-R Total Recall: β Interaction*	P-value	BVMT-R Delayed Recall: β Interaction	P-value
Hippocampus	Volume	0.0002	0.548	−.0001	0.405
Entorhinal WM	FA	−7.547	0.102	−2.450	0.188
	MD	5,523	0.021†	2,868	0.004†
Parahippocampal WM	FA	−1.788	0.705	−3.378	0.067
	MD	3,504	0.262	1,450	0.213
Temporal Pole WM	FA	−4.166	0.380	−4.350	0.024†
	MD	1,955	0.424	1,905	0.087
Left-hemispheric Structure	Imaging Biomarker	HVLT-R Total Recall: β Interaction	P-value	HVLT-R Delayed Recall:	P-value
Hippocampus	Volume	−0.0001	0.730	−0.00001	0.467
Entorhinal WM	FA	2.411	0.599	−1.095	0.641
	MD	408.2	0.923	808.0	0.690
Parahippocampal WM	FA	1.423	0.773	1.225	0.629
	MD	−7,336	0.089	−2,916	0.197
Temporal Pole WM	FA	0.4000	0.939	−0.075	0.976
	MD	−4,782	0.174	−3,029	0.098

^*

Where the interaction coefficient is β₃ in the following linear mixed effects model:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}=({\beta }_0{+\mathrm{b}}_{0\mathrm{i}})+({\beta }_1{\text{+b}}_{1\mathrm{i}}){\mathrm{Month}}_{\mathrm{j}}+{\beta }_2{\mathrm{Imaging}+\beta }_3\mathrm{Imaging}\times {\mathrm{Month}}_{\mathrm{j}}{+\mathrm{e}}_{\mathrm{ij}}$$

where b_0i = subject-specific random intercept, b_1i = random slope, and e_ij = subject error. The β₃ interaction coefficient has the units: volume, points/(month*mm³); FA, points/month; MD, points/(month*mm²/s)

^†P-value significant at alpha = 0.05

There were no significant interactions between left hemispheric WM and time. For right hemispheric WM, change in right temporal pole FA (β₃= −4.350 points/month, P=.024) and entorhinal MD (β₃= 2,868 points/[month*mm²/s], P=.004) over the 12-month study period were associated with BVMT-R Delayed Recall performance. Entorhinal MD was also significantly associated with BVMT-R Total Recall (β₃= 5,523 points/[month*mm²/s], P=.021).

DISCUSSION

In this prospective, longitudinal study of brain tumor patients undergoing RT, we demonstrate significant subacute decline in visuospatial, but not verbal, memory performance. To our knowledge, this is the first study to characterize the clinical and diffusion imaging predictors of longitudinal visuospatial memory, an important cognitive endpoint, in adult patients with primary brain tumors undergoing RT. Specifically, we found concurrent chemotherapy and the presence of seizures to be associated with greater decline in visuospatial memory at 6 months post-RT. Imaging biomarkers of damage to memory-associated gray and superficial WM structures in the right medial temporal lobe were significantly associated with poorer visuospatial memory performance. Time-dependent diffusion parameters indicating WM damage throughout the follow-up period also predicted visuospatial memory performance over 12 months.

Accounting for practice effects, verbal and visuospatial memory performance generally declined at 6 months post-RT, with a significant decline in visuospatial memory. This finding is consistent with previous studies noting impairment in verbal and visuospatial memory post-RT, in addition to decline in executive function, attention, and problem-solving²⁹. Visuospatial memory, specifically, has not yet been prospectively investigated as a correlate of diffusion imaging changes in adult patients with primary brain tumors undergoing RT, although it has been evaluated among other patient populations (i.e., small cell lung cancer patients undergoing prophylactic cranial irradiation³⁰, pediatric patients with brain malignancies undergoing RT³¹, and patients with mild cognitive impairment³² and essential tremor³³). We also found that older age was correlated with less decline in verbal memory, which differs from prior studies showing better performance of younger patients (<65 years) on both HVLT-R Total and Delayed Recall at 8 months after whole-brain radiation therapy³⁴. The reason for this unexpected finding is unclear and cannot be explained by a higher functioning elderly cohort at baseline, thus it may reflect a greater cognitive reserve within our cohort of older patients. Seizures were associated with greater decline in visuospatial memory, aligning with previous literature on worse visuospatial performance in patients with several types of epilepsy³⁵. Concurrent chemotherapy was associated with greater subacute effects on visuospatial memory, consistent with prior work showing that 30% of patients undergoing chemotherapy in addition to RT for brain tumors showed cognitive declines in visuospatial memory as well as verbal learning and memory, executive functioning, and processing speed³⁶. This relationship is likely due to a shared mechanism for brain injury and cognitive decline between chemotherapy and radiation: interference of neural stem and precursor cell function³⁷.

Across all time points over 12 months post-RT, we demonstrated that damage to right medial temporal lobe gray and WM structures were associated with poorer visuospatial memory performance. Specifically, smaller right hippocampal volumes were associated with poorer visuospatial memory performance. The association between hippocampal volume loss and memory decline has been found in several patient populations, including Alzheimer disease³⁸, temporal pole epilepsy³⁹, and traumatic brain injury⁴⁰. To our knowledge, this is the first study to show association between post-RT hippocampal atrophy and decline in memory performance in brain tumor patients on a clinical trial.

Our finding of post-RT hippocampal atrophy is also supported by prior RT-specific work, such as hippocampal dose correlations with memory decline at 6 months post-RT⁴¹ and 18 months post-fractionated stereotactic RT⁴² and functional preservation in hippocampal-sparing WBRT at 4 months post-RT⁴³. Despite the high frequency of hippocampal atrophy found in our cohort, the mean doses to the hippocampi in our cohort were much lower than previous studies⁴⁴, likely due to our study’s strict inclusion criteria for hippocampal ROIs (censoring hippocampal regions with edema or proximity to the radiated surgical cavity). Nevertheless, our data suggest an association between radiation-associated hippocampal damage and the impairment of memory performance after brain RT.

Diffusion biomarkers of white matter injury in the right medial temporal lobe were also associated with visuospatial memory performance. Radiation damage to WM tracts is thought to be caused by demyelination, axonal injury, neuroinflammation, and vascular permeability, which can result in changes in diffusion properties such as decreased FA or increased MD⁴⁵. Previous work has shown strong, dose-dependent diffusion changes post-RT in the fornix, cingulum bundle, and body of the corpus callosum⁵. We found that microstructural changes in the right entorhinal and temporal pole superficial WM are also associated with visuospatial memory performance. While previous studies have focused on tract-based analysis^5,6,10, we demonstrate the importance of WM directly beneath the cortex (i.e., superficial WM) as it enables communication across neighboring gyri in the form of U-fibers and may play a critical role in memory⁴⁶. Although prior work and clinical trials have focused on hippocampal-sparing for cognitive preservation, damage to related afferent or efferent WM pathways may also contribute to radiation-induced memory impairment^47–49, thus understanding the relationship between the sensitivity of structures within the medial temporal lobe network to radiation may prove critical to strategies for improving brain tumor treatment.

Our investigation also demonstrated significant interactions between several right medial temporal lobe WM regions and time, indicating a change in the effect of diffusion imaging parameters on visuospatial memory over time post-RT. This relationship between higher right entorhinal MD and nonverbal memory performance at later time points may reflect the more chronic, progressive damage seen later (>6 months) in the post-RT chronology of radiation-induced tissue injury^1,50. This “late delayed” brain injury is characterized by vascular abnormalities, demyelination, and even WM necrosis, which would be consistent with our findings of increased MD in the temporal pole WM^51,52. Interestingly, the interaction between right-medial temporal pole FA and time showed the opposite directionality, which may be explained by dynamic changes in WM integrity longitudinally after RT, both chronologically and biologically. Indeed, subacute (~4–6 months post-RT) brain tissue damage (i.e. transient demyelination) can be reversible¹. Thus, the counterintuitive direction of the association observed here may indicate more complex biologic processes occurring at later time points that defy our classic (and possibly simplistic) understanding of diffusion parameters. Previous work has reported an increase in FA post-RT in certain regions, which could be attributed to other, partially reversible biological processes involved in damage, such as undetected resolution of edema⁵³, axonal swelling as seen in traumatic brain injury⁵⁴, compression of peri-tumoral WM due to mass effect⁵⁵, or astrogliosis with compaction of axonal neurofilaments⁵⁶. In addition, we may not expect to see consistent evidence of radiation-induced damage across all regions, since studies suggest that WM changes are not uniform for a given RT dose distribution^8,10,57.

This work has several limitations. The neuroanatomic atlas used to auto-segment ROIs with FreeSurfer was developed based on normal brain anatomy. However, this software is robust and well-validated, not subject to manual contouring differences, is used in other patient populations with neurological disorders^32,39,40, and has been used in several other published studies of brain tumor patients^58,59. In addition, to minimize any potential segmentation error, all segmented images were carefully inspected slice-byslice, and we manually identified and censored areas of edema, tumor, and surgical cavities from all analyses, similar to previous studies^5,6,12. As discussed, there was a greater censoring of hippocampi and WM regions receiving the highest dose due to potential structural changes from nearby tumor infiltration. Thus, we likely excluded tissue from analysis where we may have found stronger correlations between imaging biomarkers of damage and memory decline, yet the associations we did find are more robust and less likely to be due to tumor or edema-related processes. Also, given the complexity of verbal and visuospatial memory outcomes, other variables (e.g., co-morbid depression or anxiety) could have influenced memory performance and future analyses can incorporate such additional variables. Finally, while our sample size is limited, we present prospective results derived from robust neurocognitive testing and detailed, consistent neuroimaging of a relatively homogenous sample of brain tumor patients, which is rare in this realm of research. Some previous studies⁴¹ combine several brain RT patient populations, including those who have received whole brain RT and partial brain RT, those with brain metastases and primary brain tumor patients. Though these limitations are worth noting, this prospective study with both detailed memory and imaging measures shows important associations between domain-specific neuroimaging biomarkers and cognitive performance after brain RT.

CONCLUSIONS

Using advanced neuroimaging techniques, we found associations between imaging biomarkers and memory performance in patients with a primary brain tumor undergoing fractionated partial RT. Concurrent chemotherapy was associated with greater decline in visuospatial memory at 6 months post-RT. Reduced hippocampal volume, decreased entorhinal FA, and greater temporal pole MD predicted worse visuospatial memory performance. Longitudinal changes in WM diffusion predicted both verbal and visuospatial memory outcomes. These findings have clinical implications, indicating that memory preservation, particularly visuospatial memory, is reliant on a variety of both gray and WM regions.

The quantitative, domain-specific data acquired through these studies will improve our understanding of brain toxicity and cognitive decline associated with radiation dosage to non-targeted tissue and can provide the basis for evidence-based cognition-sparing brain radiotherapy. Interestingly, this study introduces an association between certain WM diffusion changes and radiation-induced memory decline, which may indicate that there are other ROIs not studied in this paper that should be investigated as potential vulnerable areas contributing to post-RT cognitive decline. Further research is needed to investigate the dynamic trajectories of tissue response to radiation to better understand how MRI changes can be used to predict important neurocognitive trajectories post-treatment. Specifically, we must work to validate these associations by investigating how early imaging changes can act as biomarkers predicting subsequent memory decline on a per-patient basis. This research may support future work investigating dose-sparing protocols to avoid regions critical for memory during brain RT.

Supplementary Material

1

eFigure 1

Axial slices (inferior to superior from left to right) of T1-weighted pre-contrast MRIs for an example patient with overlayed contours of each region of interest.

Color key: red = left hippocampus; dark blue = right hippocampus; aqua blue = left entorhinal; green = right entorhinal; pink = left parahippoccampal; yellow = right parahippocampal; orange = left temporal; purple = right temporal

eFigure 2

Consort diagram for enrolled subjects with complete (N=22) and incomplete (N=34) data for baseline and 6 month assessments. “Incomplete” indicates that the subject either missed a neurocognitive assessment or MRI at either baseline or 6 months for a variety of reasons as detailed in the figure.

*11 patients declined to schedule the 6-month neurocognitive assessment. All of these patients had completed the baseline neurocognitive assessment, and 4 had completed both the baseline and 3-month neurocognitive assessment. All patients cited the reason that the neurocognitive assessments were time-consuming and they could not fit this into their schedule.

^†After initially enrolling in the study, 6 patients decided they were not interested in the study and withdrew.

^††Two patients enrolled on study, completed baseline evaluations, but then decided against completing brain radiotherapy, so were excluded from the analyses.

^§1 patient had completed all baseline and 3-month assessments, completed the 6 month neurocognitive assessment, but developed an aversion to MRIs. This patient declined the 6-month MRI and proceeded with MRIs under anesthesia on a different MRI scanner with a different MRI protocol.

eFigure 3

Spaghetti plots demonstrating each subject’s neurocognitive performance over time. The superimposed dark line connects time-specific means. The main effect of time (β₁) using the mixed model as described in the manuscript:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}={\beta }_0+{\beta }_1\times {\mathrm{Month}}_{\mathrm{j}}+{\beta }_2\times {\mathrm{Imaging}+\mathrm{u}}_{\mathrm{i}}{+\mathrm{e}}_{\mathrm{ij}}$$

was significantly associated with HVLT Delayed Recall. Specifically, neurocognitive performance significantly worsened from 3 to 12 months when accounting for all left-sided imaging parameters, in addition to worsening from 6 to 12 months when accounting for all except left hippocampal volume. HVLT Delayed Recall also significantly declined over time (β₁= −0.15, P=0.005) when evaluating time without accounting for imaging parameters as shown in the following model:

$${\mathrm{Memory}\mathrm{Scores}}_{\mathrm{ij}}={\beta }_0+{\beta }_1\times {\mathrm{Month}}_{\mathrm{j}}+{\mathrm{u}}_{\mathrm{i}}{+\mathrm{e}}_{\mathrm{ij}}$$