Neuromelanin-Sensitive MRI as Candidate Marker for Treatment Resistance in First-Episode Schizophrenia

Abstract

Objective:

Markers for treatment resistance in schizophrenia are needed to reduce delays in effective treatment. Nigrostriatal hyperdopaminergic function plays a critical role in the pathology of schizophrenia, yet antipsychotic nonresponders do not show increased dopamine function. Neuromelanin-sensitive MRI (NM-MRI), which indirectly measures dopamine function in the substantia nigra, has potential as a noninvasive marker for nonresponders. Increased NM-MRI signal has been shown in psychosis, but has not yet been assessed in nonresponders. In this study, the authors investigated whether nonresponders show lower NM-MRI signal than responders.

Methods:

NM-MRI scans were acquired in 79 patients with first-episode psychosis and 20 matched healthy control subjects. Treatment response was assessed at a 6-month follow-up. An a priori voxel-wise analysis within the substantia nigra tested the relation between NM-MRI signal and treatment response in patients.

Results:

Fifteen patients were classified as nonresponders and 47 patients as responders. Seventeen patients were excluded, primarily because of medication nonadherence or change in diagnosis. Voxel-wise analysis revealed 297 significant voxels in the ventral tier of the substantia nigra that were negatively associated with treatment response. Nonresponders and healthy control subjects had significantly lower NM-MRI signal than responders. Receiver operating characteristic curve analysis showed that NM-MRI signal separated nonresponders with areas under the curve between 0.62 and 0.85. In addition, NM-MRI signal in patients did not change over 6 months.

Conclusions:

These findings provide further evidence for dopaminergic differences between medication responders and nonresponders and support the potential of NM-MRI as a clinically applicable marker for treatment resistance in schizophrenia.

Treatment resistance in schizophrenia is a major clinical problem, with 20%–35% of psychotic patients not responding to at least two adequate trials with first-line antipsychotic treatment (1). In clinical practice, patients are often on ineffective treatment for months to years before switching to potentially effective treatment (2). This delay results in prolonged experience of side effects of ineffective antipsychotics, poorer prognosis, and decreased treatment responsiveness (2, 3). Hence, there is an urgent need for markers to identify treatment nonresponders in schizophrenia at an early stage and facilitate timely initiation of clozapine, the only antipsychotic with proven efficacy in nonresponders (4). Ideally, these markers would be noninvasive, to facilitate incorporation into clinical practice.

A well-established finding in schizophrenia is increased striatal dopamine function, which is associated with positive symptoms (5, 6). On average, patients with schizophrenia show a moderate increase in striatal dopamine synthesis capacity compared with healthy control subjects, as measured with [¹⁸F]fluorodopa ([¹⁸F]F-DOPA) positron emission tomography (PET) (7). Crucially, treatment nonresponders do not show this elevation in striatal dopamine synthesis capacity and instead show levels comparable to those of healthy control subjects (8–10). These findings suggest a different neurobiological mechanism underlying symptoms in nonresponders, which could explain why antipsychotics, predominantly acting as dopamine antagonists, are ineffective in this population (11). Moreover, they suggest the potential of dopaminergic-function markers for treatment resistance. Indeed, striatal dopamine synthesis capacity has been shown to be able to identify 40%–60% of nonresponders with a specificity of 100% and areas under the curve (AUCs) of 0.78–0.88 using [¹⁸F]F-DOPA PET (12). However, PET imaging is not feasible for routine clinical screening to identify nonresponders, given its costs, invasiveness, and associated radiation exposure.

Recently, neuromelanin-sensitive MRI (NM-MRI) has been developed to measure dopamine function noninvasively (13). Neuromelanin deposition depends on cytosolic dopamine excess in dopamine cells; it increases with dopamine synthesis (14), and it accumulates in dopaminergic cell bodies in the substantia nigra (SN), in particular the pars compacta, and in the ventral tegmental area (15, 16). These midbrain dopaminergic neurons are an anatomically heterogeneous group of cells most prominently located ventrally, in the SN pars compacta, and project predominantly to the dorsal striatum through the nigrostriatal pathway (17).

The paramagnetic neuromelanin-iron complexes that lead to T₁ reduction, in combination with reduced magnetization-transfer effects, result in high signal intensity in the SN using NM-MRI (18, 19). NM-MRI signal correlates with regional neuromelanin concentration in postmortem SN tissue and to PET measures of dopamine release in the dorsal striatum in vivo (20). Furthermore, NM-MRI exhibits excellent test-retest reliability, including for voxel-wise analyses (20–22). SN NM-MRI is therefore a reliable proxy measure of dopamine function in the nigrostriatal pathway. In schizophrenia patients, NM-MRI SN signal is increased compared with healthy control subjects (23). In particular, signal in a ventral/anterior subregion of the SN has been found to be positively associated with positive symptoms (20). These results are in line with previous PET studies on presynaptic dopamine function in the striatum and SN (7, 24) and support NM-MRI’s ability to index hyperdopaminergic function in schizophrenia. However, NM-MRI has not yet been investigated in treatment-resistant schizophrenia. The aim of the present study was to assess whether, in line with earlier [¹⁸F]F-DOPA PET results, NM-MRI SN signal is elevated in treatment responders compared with nonresponders, a crucial first step to evaluating NM-MRI’s potential as an noninvasive marker for treatment resistance in first-episode schizophrenia.

METHODS

Participants

We recruited first-episode psychosis patients with a primary diagnosis in the schizophrenia spectrum through specialized early psychosis clinics in Amsterdam. All patients were diagnosed according to DSM-5 criteria by a specialized early psychosis psychiatrist and were undergoing treatment at a specialized early psychosis clinic. Healthy volunteers were recruited through advertisements as a control group. All participants were between 18 and 35 years of age. Exclusion criteria included use of antipsychotic medication longer than 1 year; past or present substance use disorder and current substance use other than nicotine, alcohol, and cannabis; use of dopaminergic drugs; neurological disorder or brain damage; MRI contraindications; and pregnancy. We allowed for medication use up to 1 year and cannabis use to acquire a representative group of first-episode patients with an initial referral to a specialized early psychosis clinic. Additional exclusion criteria for control subjects included current or past diagnosis of a psychiatric disorder, psychotropic medication use, and a positive family history (first- and second-degree relatives) for psychotic disorders. Control subjects were matched on age, gender, IQ, and smoking status. The study was approved by the Medical Ethics Committee of the Amsterdam Medical Center, University of Amsterdam (METC 2017_307) and registered at the Central Committee on Research Involving Human Subjects (NL63410.018.17). Before study entry, participants provided written informed consent after receiving a complete description of the study.

Study Design

All participants were assessed at baseline, and those in the patient group were followed up at 6 months to determine treatment response status. Baseline measurements included an MRI scan, a clinical interview, and an IQ test. During the interview, general information was collected, including demographic characteristics, current and past medication use, and medical history. Symptom severity in patients was measured using the Positive and Negative Syndrome Scale (PANSS) (25). The Mini-International Neuropsychiatric Interview was administered in the healthy control subjects to rule out the presence of any psychiatric disorder (26). Global functioning was measured using the Global Assessment of Functioning Scale (GAF). Alcohol and drug use was measured with the Composite International Diagnostic Interview (CIDI) (27). IQ was determined by the shortened version of the Wechsler Adult Intelligence Scale (WAIS-III) (28). Socioeconomic status was determined based on parental education and occupation (29). In addition, if consent was given, antipsychotic drug serum levels were measured, and/or patients’ relatives were contacted to evaluate adherence to antipsychotic medication. Patients were excluded from analysis if treatment response could not be determined because of antipsychotic drug serum levels that were below the therapeutic range, as defined by ARUP laboratories (https://ltd.aruplab.com), or if patients were nonadherent according their relatives. During the entire study, patients had prescriptions for antipsychotic treatment, based on standard clinical guidelines, from their treating psychiatrist.

Treatment response was assessed after 6 months in a follow-up clinical interview. Patients were considered treatment nonresponders if they showed nonresponse despite at least two trials of different antipsychotics with a duration of at least 6 weeks, or after 4 weeks if there was complete absence of response or if severe side effects occurred, at adequate antipsychotic dosages for first-episode psychosis as defined by the Dutch Multidisciplinary Guidelines for Schizophrenia (30). Nonresponse was defined when at least one of the following items of the PANSS was scored as moderate or higher (score ≥4): delusions (P1), conceptual disorganization (P2), hallucinatory behavior (P3), mannerisms and posturing (G5), and unusual thought content (G9) (31). Patients were also classified as nonresponders if their medication was switched to clozapine during the follow-up period. For patients who were lost to follow-up, response was determined based on medical files and contact with their clinician. Patients were classified as nonresponders if they were treated with clozapine and as responders if the clinician reported a clear and sustained overall improvement in symptoms on standard antipsychotics. If treatment response could not be determined, patients were excluded. If consent was given, a follow-up blood draw was conducted to measure antipsychotic serum levels, and patients’ relatives were contacted again to evaluate adherence to antipsychotic medication. Additionally, a follow-up MRI scan was conducted in patients who consented.

NM-MRI Acquisition

All MRI scans were acquired using a 3-T scanner (Philips, Ingenia Elition X; Best, the Netherlands) with a 32-channel head coil. Participants were asked to refrain from alcohol and cannabis use for 24 hours in advance of the scan. Prior to scanning, urine drug screening was performed, as well as a pregnancy test for women. For NM-MRI slice placement and preprocessing, whole-brain structural T¹-weighted volumetric images were acquired (TE=4.1 ms, TR=9.0 ms, 189 slices, FOV=284×284×170 mm, voxel size=0.9×0.9×0.9 mm, flip angle=8°). For NM-MRI images, a two-dimensional gradient recalled echo sequence with off-resonance magnetization transfer pulses was performed (TE=3.9 ms, TR=260 ms, 8 slices, FOV=162×199 mm, slice thickness=2.5 mm, flip angle=40°, resolution=0.39×0.39 mm², number of signal averages=2, magnetization transfer frequency offset and duration, 1200 Hz and 15.6 ms). The NM-MRI was oriented perpendicular to the fourth ventricle floor, with slices placed from the posterior commissure to halfway through the pons.

NM-MRI Analysis

The NM-MRI was preprocessed using a previously validated pipeline (22). The pipeline was written in MATLAB and included ANTs (32), SPM12 (33), and AFNI (34) routines. In short, the preprocessing steps were as follows: 1) NM images were coregistered to the T₁-weighted image using ANTs Rigid registration; 2) brain extraction was performed for the T₁-weighted images using ANTs Brain Extraction; 3) T₁-weighted images were spatially normalized into MNI standard space using ANTs MNI deformable registration; 4) the coregistered NM-MRI images were spatially normalized by applying the warping parameters that were used to normalize the T₁-weighted images to MNI space; and 5) the normalized NM-MRI images were spatially smoothed with a 1-mm full width at half maximum Gaussian kernel using AFNI smoothing. After preprocessing, all images were visually inspected using standardized quality-control summary outputs (an example is provided in the online supplement).

For the a priori voxel-wise analyses of the NM-MRI signal in the SN, a mask of the SN-ventral tegmental area (SN-VTA) complex was used. This mask was created in a previous study by manually tracing the hyperintense region representing the SN-VTA complex of an NM-MRI template (20). The tracing was deliberately overinclusive to facilitate inclusion of all voxels of the SN-VTA complex from all participants. The neighboring hypointense region constituting the crus cerebri was traced as a reference region since it is known to have negligible NM content (20).

The NM-MRI signal from these regions was used to calculate the contrast ratio (CR). The CR at each voxel was calculated as percent NM-MRI signal difference between each voxel in the SN-VTA mask (I_V) and the mode of the signal intensity in the crus cerebri (I_CC), with the following formula:

$${\mathit{CR}}_V=\left\{\frac{[{\text{I}}_{\text{V}}-\text{mode}({\text{I}}_{\text{CC}})]}{\text{mode}({\text{I}}_{\text{CC}})}\right\}*100$$

The mode I_CC was calculated for each participant from a kernel-smoothing-function fit of a histogram of all voxels in the crus cerebri mask, as this enhances robustness to outliers and edge artifacts (20).

The a priori analysis tested differences between nonresponders and responders using a voxel-wise robust linear regression predicting CR within the SN-VTA based on treatment response status, with age as covariate (given age effects in NM concentration and NM-MRI) (35). Missing values because of incomplete SN coverage or extreme values (smaller or greater than the 1st and 99th percentiles; CR< −0.11 and CR>26.67) were excluded (on average, 39 voxels [SD=32], or 2.2% of the SN-VTA voxels per patient). In keeping with previous work, we chose voxel-wise analyses over region-of-interest analyses to reduce statistical circularity in defining regions via signal intensity thresholding and to account for regional heterogeneity of dopamine neurons across tiers without well-defined anatomical boundaries. Hypothesis testing was based on a permutation test (10,000 permutations) for the treatment response status variable, which determined the chance-level distribution of the number of SN-VTA voxels exceeding a threshold of p<0.05 (see the Supplementary Methods section in the online supplement). To obtain a measure of effect size unbiased by voxel selection, a leave-one-subject-out analysis was performed. Here, significant voxels for each patient were identified in a voxel-wise analysis that included the complete patient sample except the left-out patient. The significant voxels were then used to extract the mean CR for this left-out patient.

Post hoc receiver operating curve (ROC) analysis was performed to assess the ability of NM-MRI to identify nonresponders without mislabeling the responders by extracting the AUC and the sensitivity at 100% specificity. Treatment response status (responder/nonresponder) was used as outcome variable and CR as independent variable. For completeness, we also show the results of the region-of-interest analysis using the whole SN-VTA (see the Supplementary Methods section in the online supplement). In addition, the mean CR_SN-VTA from the region-of-interest analysis was compared between nonresponders and responders using a one-way analysis of covariance (ANCOVA) with age as covariate.

For post hoc analyses including the control group, the mean CR from the significant voxels (thresholded at p<0.05 and surviving the permutation-based family-wise-error correction for the a priori voxel-wise analysis) and the mean CR_SN-VTA were compared between nonresponders, responders, and control subjects using a one-way ANCOVA with age as covariate.

A secondary aim was to assess the robustness of NM-MRI over 6 months, by comparing the mean CR for all patients between baseline and follow-up. A linear mixed-effect model was conducted with time (baseline, follow-up) as the independent variable while controlling for treatment response status and including subject as a random effect.

Exploratory correlations assessed the relationship of the mean CR of significant voxels to clinical variables. Additional exploratory post hoc ANCOVAs were performed, controlling for clinical variables and attrition (see the Supplementary Results section in the online supplement).

In general, appropriate parametric tests were used (or nonparametric tests when normality assumptions were violated), and results of post hoc analyses were considered statistically significant at a p threshold of 0.05.

RESULTS

The baseline demographic and clinical characteristics of the participants are summarized in Table 1 (for diagnoses and medication use, see Table S1 in the online supplement). In total, 79 patients and 20 control subjects were included in the study. After exclusions (see the Supplementary Results section and Table S2 in the online supplement), 62 patients and 20 control subjects were retained. A subgroup of 37 patients participated in the follow-up MRI scan (see Tables S3 and S4 in the online supplement).

TABLE 1.

Baseline demographic and clinical characteristics of treatment responders and nonresponders and healthy control subjects^a

	Responders (N=47)		Nonresponders (N=15)		Control Subjects (N=20)
Characteristic	N	%	N	%	N	%	p
Male	32	68.1	11	73.3	14	70.0	0.96
Race/ethnicity							0.20
White	27	57.4	10	66.7	18	90.0
Black	10	21.3	3	20.0	0	0.0
Latinx	2	4.3	0	0.0	1	5.0
Mixed	8	17.0	2	13.3	1	5.0
Smoker	23	48.9	6	40.0	9	45.0	0.83
Cannabis user	28	59.6	6	40.0	12	60.0	0.38
	Mean	SD	Mean	SD	Mean	SD	p
Age (years)	24.06	4.63	21.27	3.20	22.70	4.08	0.14
IQ	92.28	14.53	91.53	11.75	92.95	10.88	0.95
Socioeconomic status	21.95	10.38	21.07	10.18	27.68	8.93	0.09
Nicotine use (cigarettes/day)	6.38	7.77	5.13	7.19	7.75	9.70	0.69
Cannabis use (weeks/year)	15.34	20.18	13.53	20.97	21.1	24.25	0.49
GAF score	58.17	11.37	50.13	7.97	79.05	6.69	<0.001
Age at illness onset (years)	23.36	4.55	20.33	3.37			0.03
Illness duration (weeks)	38.85	40.79	50.90	39.78			0.12
Antipsychotic duration (weeks)	12.87	9.28	19.84	10.15			0.02
Antipsychotic dosage (mg/day, chlorpromazine equivalents)	397.33	164.63	330.83	134.04			0.16
PANSS
Positive score	10.59	3.46	14.13	5.29			0.03
Negative score	11.32	4.32	14.80	5.48			0.02
General score	22.62	4.47	26.20	6.06			0.04

^aGAF=Global Assessment of Functioning Scale; PANSS=Positive and Negative Syndrome Scale. Socioeconomic status is based on the mean socioeconomic status of the parents, calculated by Hollingshead Four-Factor Index, which combines parents’ education and occupation.

No significant differences were found between the three groups in sex, age, IQ, nicotine use, and cannabis use at baseline. The nonresponders and responders did not differ significantly in duration of illness or antipsychotic dosage. Nonresponders had a lower age at illness onset, a longer duration of antipsychotic use, lower scores on the GAF, and higher scores on all subscales of the PANSS.

NM-MRI Analyses of Treatment Response Status

Consistent with our hypothesis, the a priori NM-MRI voxel-wise analysis of treatment response status revealed a subset of significant voxels (treatment response voxels: 297 of 1,807 voxels at p<0.05, robust linear regression controlling for age; permutation-corrected p=0.032; peak voxel MNI coordinates: x=−4, y=−23, z=−20) with increased CR in responders compared with nonresponders (Figure 1). Repeating the analysis without excluding extreme voxel values yielded similar results (treatment response voxels: 307 of 1,807 voxels at p<0.05, corrected p=0.031). Treatment response had a moderate unbiased effect on NM-MRI signal in the treatment response voxels (unbiased leave-one-out Cohen’s d=0.46, 95% CI=−0.14, 1.06). The region-of-interest analysis also showed significantly higher CR_SN-VTA in the responders compared with the nonresponders (F=4.82, df=1, 59, p=0.032).

FIGURE 1. NM-MRI analyses of treatment response status in first-episode schizophrenia^a.

^aPanel A is a map of treatment response voxels, in which responders exhibited a higher neuromelanin-MRI contrast ratio (CR) than responders (thresholded at p<0.05, permutation-corrected p=0.032). In panel B, receiver operating characteristic curves display sensitivity and specificity in separating the two patient groups using the mean CR of the treatment response voxels from the voxel-wise analysis, the unbiased leave-one-out (LOO) analysis, and the region-of-interest (ROI) analysis. In panel C, scatterplots show the extracted contrast ratio from the region-of-interest analysis (CR_SN-VTA), with the box plots showing the group mean and standard deviation. A significant effect of group on mean CR_SN-VTA (p=0.02) was found.

The ROC analysis of the treatment response voxels yielded a good AUC of 0.85, although these effects are inflated as a result of circularity in voxel selection. Noncircular analyses yielded more moderate effects (unbiased leave-one-out AUC=0.62, region-of-interest-based AUC=0.68) (Figure 1).

Post hoc analyses in the treatment response voxels showed group differences such that responders had higher CR than both nonresponders and control subjects (see the Supplementary Results section in the online supplement). Because these results will be biased as a result of voxel selection for separating responders and nonresponders, we then performed a similar analysis using the mean CR in the unbiased whole SN-VTA (mean CR_SN-VTA). Here, there was a significant effect of group on the mean CR_SN-VTA (F=4.1, df=1, 78, p=0.02) (Figure 1C). Age was not a significant covariate (F=0.12, df=1, 78, p=0.73). However, post hoc Tukey tests showed differences between responders (N=47; mean=13.79, SD=1.24) and both nonresponders (N=15; mean=12.99, SD=1.14) and healthy control subjects (N=20; mean=13.06, SD=1.06) that fell short of statistical significance (p=0.07 and p=0.09, respectively), and no significant difference between nonresponders and healthy control subjects (p=0.99).

Secondary Longitudinal Analyses

To assess differences in average CR in the treatment response voxels between baseline and follow-up, we ran a linear mixed-effect model analysis with the 62 baseline measurements and 37 follow-up measurements. There was a significant main effect of treatment response status (beta=−1.45, t=−3.77, p<0.001). There was no significant main effect of time (beta=−0.06, t=−0.26, p=0.80), nor an interaction of time by treatment response status (beta=0.89, t=1.66, p=0.10). At follow-up, the differences in mean CR between nonresponders (N=9; mean=13.47, SD=1.55) and responders (N=28; mean=15.25, SD=1.70) remained significant after controlling for age (F=7.66, df=1, 34, p=0.01) (Figure 2). Age was again not a significant covariate (F=0.47, df=1, 34, p=0.50).

FIGURE 2. Secondary longitudinal analyses^a.

^aScatterplots show the extracted neuromelanin contrast ratio from the treatment response voxels (CR_TRV) of all individual patients at baseline and follow-up, with the thick lines representing the longitudinal linear mixed-effect model. At baseline and follow-up, a significant difference was found between nonresponders and responders. No significant main effect of measurement was found.

*p<0.05. ***p<0.001.

Exploratory Symptom Correlations

Spearman’s rank correlation yielded a significant negative correlation between average CR in the treatment response voxels and positive symptoms and general symptoms at baseline, respectively (r=−0.26, df=60, p=0.041, and r=−0.25, df=60, p=0.045). This was not significant in a partial correlation controlling for group (r=−0.15, df=59, p=0.26, and r=−0.14, df=59, p=0.28, for positive and general symptoms, respectively). Nor did this correlation reach significance within the nonresponder (r=−0.36, df=13, p=0.19) or responder group (r=−0.02, df=45, p=0.89) for the positive symptoms, or for the general symptoms in nonresponders (r=−0.22, df=13, p=0.43) or responders (r=−0.09, df=45, p=0.55). No significant correlations were found between average CR in the treatment response voxels and age (r=0.03, df=60, p=0.78), illness duration (r=−0.06, df=60, p=0.62), medication duration (r=0.01, df=60, p=0.95), medication dosage (r=−0.02, df=60, p=0.87), and negative symptoms (r=−0.19, df=60, p=0.15).

DISCUSSION

To our knowledge, this is the first study evaluating the potential of NM-MRI as a noninvasive marker for treatment resistance in first-episode schizophrenia. In line with our hypothesis, nonresponders showed significantly lower NM-MRI signal compared with responders, and similar NM-MRI signal compared with control subjects. NM-MRI is able to identify nonresponders with AUCs of 0.62–0.85. The significant voxels associated with response status were predominantly localized in the ventral tier of the SN. Furthermore, NM-MRI appears to be relatively robust in that NM-MRI signal remained stable over 6 months of follow-up and was uncorrelated with illness duration, medication duration, or medication dosage.

Our results are consistent with the finding of lower striatal dopamine synthesis capacity using [¹⁸F]F-DOPA PET in nonresponders compared with responders (8–10) and provide further evidence that nonresponders differ from responders in nigrostriatal dopaminergic functioning. NM-MRI is a less direct measurement of dopamine function than [¹⁸F]F-DOPA PET. [¹⁸F]F-DOPA PET captures state-dependent changes in dopamine functioning, whereas NM-MRI appears to be a more stable, trait-like measure, likely owing to the slow timescale of neuromelanin accumulation. In addition, factors other than neuromelanin concentration alone may contribute to the NM-MRI contrast, including myelination (36), and further work is needed to elucidate the specificity of the contrast to neuromelanin concentration. However, NM-MRI has several advantages over PET imaging, including lower costs, noninvasiveness, and no ionizing radiation, all factors that are essential for a clinically applicable marker to identify treatment resistance in schizophrenia.

The subset of voxels significantly associated with response status were mainly localized to ventral SN. The ventral SN provides dopaminergic innervation of the associative striatum (37). Previous studies have established that presynaptic dopamine function in schizophrenia is especially elevated in the associative striatum (38). A recent NM-MRI study found that voxels in the SN associated with psychosis severity in unmedicated patients also predominated in the ventral (and anterior) aspects of the SN (20). Together, these results support a specific involvement of the ventral SN in the pathophysiology of psychosis in schizophrenia, highlighting the functional importance of SN topography.

This study employed a naturalistic design; the majority of patients were recruited from a tertiary inpatient clinic within several weeks after their admission and had a history of medication use prior to admission via general practitioner or other clinics (e.g., crisis services or secure services). Given practical and ethical reasons, it was unfeasible to include patients at illness onset or during the period when symptoms were most severe. As a result, most of the included patients showed at least partial response to antipsychotics at baseline, and the patient groups differed in symptom severity and medication duration. This could have affected our results, but we observed no change in NM-MRI signal in the treatment response voxels over a 6-month period while patients were on antipsychotic medication and their positive symptoms significantly improved (see Table S3 in the online supplement). It is still possible that NM-MRI signal change may occur over longer follow-up periods given the slow temporal dynamics of neuromelanin accumulation. Longer follow-up periods might be challenging in first-episode psychosis, though, as follow-up is generally difficult in this patient population. Indeed, we have lost 10 patients to follow-up, and a considerable number were not motivated to participate in the follow-up MRI scan session. The lack of correlations with illness duration, medication duration, and antipsychotic dosage, however, also suggests small or no effects of disease-related aspects, or of medication, on NM-MRI signal in treatment response voxels during early stages of first-episode psychosis. We also found no significant effect of age on NM-MRI signal in these voxels within this young age range (35), consistent with a meta-analysis of NM-MRI in schizophrenia (23). Collectively, these results further support NM-MRI as a candidate marker for treatment resistance in first-episode psychosis in a naturalistic setting, where it may be unfeasible to limit markers to patients scanned at illness onset or before medication use. Circumstances may vary between countries or clinics, but it is not uncommon for patients to receive crisis treatment before being referred to specialized care, where MRI scans can be obtained. An MRI-based marker in such settings (i.e., after start of antipsychotic medication) should still be able to aid in therapeutic decisions and have utility in reducing delays in receiving effective treatment.

It is important to mention that the majority of the non-responders in our study, although meeting accepted criteria, were not completely treatment resistant. They experienced some effect of standard antipsychotics without reaching remission of positive symptoms. The significant negative correlation between symptom severity and NM-MRI signal did not hold when controlling for group, and it is therefore likely driven by nonresponders (with more severe positive symptoms) having lower NM-MRI signal. It may also, however, hint at the notion that approaching response as a continuum is more appropriate than a binary division, although the latter may facilitate clinical decision making (11, 39). A study evaluating response dimensionally, via changes before versus after treatment initiation, would be necessary to assess this notion and could account for some of the overlap in the NM-MRI signal, assuming signal scales with response magnitude. In addition, results might differ for ultra-treatment-resistant patients (non–clozapine responders). In our sample, 12 of the 15 nonresponders responded to clozapine at follow-up. The relatively large number of nonresponding patients who received clozapine treatment is likely due to the fact the majority of our sample were treated by a specialized early psychosis group with extensive experience in clozapine treatment. Furthermore, the fact that the majority of the nonresponders responded to clozapine could be a result of the early initiation of clozapine, as this can increase clozapine responsiveness (3, 40).

In summary, this study demonstrates the potential of NM-MRI as a noninvasive marker for treatment resistance in schizophrenia at an early stage and provides further evidence for heterogeneity in the neurobiology of schizophrenia. Further research is needed to determine the out-of-sample predictive value of NM-MRI for nonresponders in larger samples, for example, by applying machine learning approaches. Eventually, an adequate prediction model could lead to early identification of treatment resistance in schizophrenia and thereby substantially reduce delays in effective treatment and improve outcomes.

Data availability:

The analysis pipeline (MATLAB scripts) and treatment response voxel mask are available upon request from the authors.