Personalized magnetic brain stimulation for treatment-resistant depression, meaning depression that has not responded to standard antidepressants, works by finding a spot on the left frontal cortex. That spot is chosen because its activity at rest moves most opposite to a small deep region, the subgenual anterior cingulate cortex (sACC). The method's output is a coil coordinate. The input is a decision about where the sACC begins and ends. That boundary is ambiguous, and expert annotators disagree about it. Nobody has measured how far this disagreement moves the final coil target, or whether that movement is large next to the other sources of error in the pipeline. The biggest of those other sources is the noise of the resting-state brain scan itself.
We propose to measure this directly and then act on it. We will collect multiple expert delineations of the sACC on depressed patients' scans and train a model that reproduces their disagreement. We will push samples from that model through the targeting pipeline and split the spread of the resulting coil targets into two parts: the part caused by seed-boundary ambiguity and the part caused by scan-to-scan noise. Then we choose the target that performs best on average across both sources of uncertainty, rather than against a single fixed seed. The core test is whether this "robust target" holds steadier across repeat scans, and engages the intended network more reliably, than the current deterministic approach. If seed ambiguity turns out to be negligible, that is a useful negative result: the field could stop worrying about seed definition and spend its effort on scan length instead.
Core Claim
The current personalized protocol, TAO-TMS, is from Kong et al.. It finds the left dorsolateral prefrontal location most anti-correlated with the sACC and delivers accelerated theta-burst stimulation there. The authors report a 70% response rate in 20 naturalistic patients, a 21% gain in electric-field network engagement over the scalp-based BeamF3 placement, and a mean offset of 13.4 mm from BeamF3. Their pipeline treats the sACC seed as fixed ground truth.
Our claim is twofold. First, the target is a random variable with two separable sources of variance, seed boundary and functional-data noise. Second, a target chosen to maximize expected anti-correlation under both sources is more reproducible than one chosen from a single seed and a single scan. The scale that decides whether any of this matters comes from Opitz et al.. In healthy subjects, they showed that stimulation engages two to three cortical patches spaced 7.7 ± 2.3 mm apart, and that center-of-mass estimates miss fitted targets by 4.6 mm. Target shifts of a few millimetres are therefore functionally meaningful, while shifts well under that scale are probably not.
Research Program
The first step builds the annotation resource that does not exist today. Several trained neuroanatomists will independently delineate the bilateral sACC on T1-weighted scans of depressed patients. Each patient will also have at least two resting-state sessions, so that scan noise can be measured separately from boundary noise. Alongside the human seeds we will generate the automated seeds that practitioners actually use: a coordinate sphere in standard space, FreeSurfer-based labels, and the probabilistic cytoarchitectural labels of NextBrain. NextBrain, from Casamitjana et al., is a histology-derived atlas giving voxel-wise region probabilities. It reduced registration error to 0.99 mm and outperformed FreeSurfer on Alzheimer's classification, but it has never been tested on psychiatric targeting.
The second step learns a distribution over seeds. Here we borrow methods rather than invent them. The Latent Doctor Model of Willemink et al. is a conditional variational autoencoder: a generative model that, given an image, samples plausible annotations from a learned latent "annotator style" space. The rater-specific Bayesian multi-decoder of Hu et al. separates human rater variability from model uncertainty. Both were validated mostly in 2D, and Li et al. flag 3D memory cost as the bottleneck. The sACC, however, is small, so we can work on a cropped 3D patch around the cingulate, which keeps these models tractable. We need this learned model so that future patients can receive a seed distribution without a panel of annotators.
The third step propagates and optimizes. Let $s$ be a seed sampled from the learned seed distribution and $b$ a bootstrap resample of the patient's resting-state time points. Let $c_b(t, s)$ be the correlation between the time series at candidate cortical location $t$ and the mean time series in seed $s$. The robust target is
$$ t^{*} = \arg\max_{t} \; \mathbb{E}_{s,b}\left[-c_b(t,s)\right], $$
with a lower-quantile version as a risk-averse variant. We then decompose the variance of the per-sample argmax target location into a seed component and a data component using the law of total variance. Every candidate target then goes through realistic head-model electric-field simulation, so that we score network engagement and not only geometric distance.
This differs from the diagnosis-first approach of Wu et al.. That work weights rater masks by their usefulness to a downstream classifier and gains up to 2.73% in diagnostic AUC on 2D images. Our downstream utility is cross-session target reliability rather than clinical outcome. This substitution is deliberate. With cohorts of about 20 patients, training seed weights on response labels would overfit, whereas reliability can be measured on every scan pair. We set aside deep brain stimulation tract delineation, despite the open questions raised by Desai et al. and Alagapan et al.. Their 10-patient cohorts cannot support a variance decomposition.
Baselines
The naive reference point is BeamF3 scalp placement, which ignores both the seed and the scan. Kong et al. report a 21% response rate for a historical BeamF3 cohort, against 70% for TAO-TMS. That response comparison involves different schedules and is not a clean comparison of targeting alone. We will therefore compare BeamF3 to our method on engagement and stability within the same subjects, not on response.
The strongest existing method is deterministic TAO-TMS-style targeting from a single atlas sACC seed on a single session. Its reported benchmarks are the 70% response rate and the 21% engagement gain over BeamF3 from Kong et al. We found no reported result on Lacuna for the cross-session reproducibility of sACC-anticorrelation targets, which is our primary metric, so we will measure it ourselves. All arms share the same scans, preprocessing, candidate cortical mask, and head models. Beating this baseline means two things on held-out sessions: smaller target displacement between sessions, and higher electric-field engagement of the network defined by the other session's data. The improvement must exceed paired-bootstrap confidence bounds.
Three ablations isolate the key ingredient. A data-bootstrap-only target uses the single atlas seed but averages over time-point resamples. A seed-only target uses the learned seed distribution on one unresampled session. A consensus-seed target uses a majority-vote or STAPLE seed with no distribution, which tests whether modeling disagreement adds anything beyond a better single mask. A NextBrain-probability-weighted seed serves as the zero-annotation competitor.
Why not just…?
Why not just bootstrap the fMRI data and ignore the seed? This is the most serious threat. Scan noise may dominate boundary noise. The data-bootstrap-only ablation tests exactly this. If the seed component of target variance is small, for example well below the 4.6 mm center-of-mass error scale from Opitz et al., we will report that seed modeling is unnecessary and recommend longer scans instead.
Why not just use a consensus mask? Willemink et al. show that majority-vote training makes models overconfident on ambiguous cases, and Wu et al. show that consensus masks are suboptimal for downstream tasks. Both findings come from pathology and 2D imaging, however. The consensus-seed ablation settles whether the same holds for a seed whose only job is to produce a mean time series. Averaging over voxels may wash out boundary differences.
Why not just use NextBrain's probability map as a weighted seed? It costs no annotators and already encodes anatomical variability across specimens. If it matches the learned rater distribution on stability and engagement, it is the better practical choice and we will say so. The rater model would then serve only to show that NextBrain captures human ambiguity. Our expectation that NextBrain could fall short rests on its being built from five post-mortem hemispheres and never validated in depressed brains, but that is a hypothesis we will test, not a known result.
Proposed Experiments and Decisive Checks
The first decisive check is the variance decomposition itself. For each patient, we report the seed and data components of target variance in millimetres. We also report the fraction of seed samples whose argmax lands on a different cortical patch, using the patch scale of Opitz et al. as the threshold for "different."
The second check is held-out reproducibility. We compute targets on session one and evaluate engagement on session two, comparing the robust target against all baselines and ablations above.
The third check is generalization of the learned seed model. We hold out entire annotators and test whether samples from the model cover their delineations, measured by generalized energy distance and the staged-threshold Dice used in QUBIQ (Li et al.). Top QUBIQ ensembles reached Dice of 0.80 to 0.85 on other organs. That serves only as a rough reference, since no sACC result exists.
Finally, if TAO-TMS-style outcome data become available through collaboration, we will run one exploratory test: whether per-patient target uncertainty, or the distance between deterministic and robust targets, relates to response. We will state in advance that a 20-patient cohort can show at most a direction, not proof.
References
- Efficacy of Personalized Connectivity-Guided Accelerated Brain Stimulation in a Naturalistic Treatment-Resistant Depression Population
- Determining the cortical target of transcranial magnetic stimulation
- A probabilistic histological atlas of the human brain for MRI segmentation
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Inter-Rater Uncertainty Quantification in Medical Image Segmentation via Rater-Specific Bayesian Neural Networks
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Longitudinal Changes in Cortical Response Dynamics with Deep Brain Stimulation to the Subcallosal Cingulate for Treatment-Resistant Depression
- Cingulate dynamics track depression recovery with deep brain stimulation