Brain segmentation networks trained mostly on healthy anatomy tend to "repair" abnormal brains. When a ventricle is grossly enlarged or a structure is displaced, the network quietly bends the output back toward normal anatomy. Retraining the network to favor pathology makes this better on sick brains and worse on healthy ones. Different scanners and sequences add another source of error that looks like biology. This proposal puts a small, auditable rule base on top of two existing segmentation models: a normative one and a pathology-weighted one. The rule base checks each model's output against anatomical rules and acquisition-aware rules that a clinician could read, and it decides per case and per structure which output to trust, or whether to abstain. It also records which rule fired and why.
The point is to resolve a documented trade-off with a component that can be inspected, not to build a better segmentation network. Shang et al. raised mean Dice on severely pathological fetal brains from 0.60 to 0.67 by weighting synthetic training toward abnormal shapes, and paid with a drop on neurotypical brains. If rule-based checks can tell when each model's output is plausible, a routed system should keep the pathological gain without the healthy-brain penalty. The fired rules would then double as a machine-checkable explanation of what went wrong.
Core Claim
Cheap, declarative checks on a segmentation's own outputs can separate three situations that a network's confidence score blurs together. The first is that the anatomy is normal and the normative model is right. The second is that the anatomy is abnormal and the normative model has erased it. The third is that the acquisition does not support a reliable measurement of a given structure.
The most diagnostic check is a label–intensity consistency rule, computed relative to each scan's own tissue intensity distribution. Voxels labelled as brain tissue whose intensity matches the case's cerebrospinal fluid are a signature of "hallucinated normalcy." Because the rule is relative to the scan, it does not depend on the scanner's arbitrary intensity scale, which Wiltgen et al. identify as a core obstacle to pooling data.
Two further families of rules complete the base. Relational and topological rules state, for example, that the ventricles are enclosed by white matter and that compartment volumes sum to intracranial volume. Acquisition rules mark which measurements a protocol can support, for example cortical thickness under thick slices or tissue contrast at 0.55T. The claim is that routing on these rules beats routing on network uncertainty or on a learned quality predictor. The advantage should be largest when the pathology type or protocol is held out from whatever the alternative was trained on, and the rules should produce explanations that experts agree with.
Relation to Prior Work
The closest work joins rules, generalization and interpretability for diagnosis. XAI-MeD runs a symbolic branch that encodes clinical guidelines next to a neural branch and routes between them with entropy-based criteria. NEURO-GUARD has a language model write executable diagnostic code from retrieved guidelines. KG-DG fuses lesion-count rules with Vision Transformer logits for diabetic retinopathy. All three use rules to reach a diagnosis, and all three take the upstream detectors' outputs as given.
KG-DG itself is the constraining evidence here: its continuous vessel-morphology features hurt cross-domain performance. That is direct evidence that extracted primitives fail under acquisition shift, and nothing in this line audits them. Our rules are therefore not diagnostic rules. They verify a network's intermediate output, conditioned on the acquisition, and they apply to 3D brain MRI rather than to 2D fundus images.
Two papers supply methods. RelateSeg relaxes first-order spatial rules into fuzzy, differentiable losses for 2D natural scenes. We borrow its fuzzy truth values, but evaluate the rules at inference time as verifiers rather than optimizing against them, so a violation remains visible instead of being optimized away. The coherence loss of Patrício et al. restricts concept activations to the lesion region. Our label–intensity rule plays an analogous grounding role for segmentation outputs. The rule engine itself is ordinary forward chaining over a separate knowledge base, the decoupling argued for in Socrates and used clinically by Woo et al..
Three papers motivate the protocol side. Physics-Informed Brain MRI Segmentation feeds sequence parameters into the network and cuts upgrade-induced volume discrepancy on SABRE to 2%. It also reports slightly worse gray-matter consistency on SPGR, the kind of sequence-specific unreliability an acquisition rule should report rather than hide. SynthSeg shows that randomizing contrast and resolution in synthetic training yields contrast-agnostic segmentation, and it is our normative model. Mamalakis et al. show that a single model learning tumor domains forgets healthy ones unless modules are explicitly frozen. That supports keeping two specialized models and choosing between them, rather than forcing one model to serve both populations.
Research Program
The system has three parts. The first is a pair of frozen segmenters: SynthSeg trained on normative fetal templates, and the pathology-weighted model of Shang et al., which uses shape-cluster sampling and synthetic ventricular dilation.
The second is a measurement layer. It computes, for each case and structure, the following quantities: volumes normalized to gestational age, adjacency and enclosure relations, left–right asymmetry, a compartment-sum residual, and the label–intensity consistency score. The consistency score is the fraction of tissue-labelled voxels whose intensity falls in the case's own cerebrospinal-fluid mode. Acquisition metadata (field strength, slice thickness, reconstruction method) comes from headers.
The third is a knowledge base of roughly 30–60 rules written with fetal neuroradiologists. Each rule carries a fuzzy truth value and resolves to one of four verdicts per structure: trusted, normative-model failure, pathology-consistent, or unverifiable under this protocol.
Routing is deterministic. If the normative output passes every rule, it is used. If it fails the consistency or enclosure rules and the pathology model's output passes them, the pathology output is used. If both fail, or an acquisition rule declares the measurement unsupported, the system abstains for that structure.
Every decision comes with the chain of rules that fired. That chain is the explanation we evaluate. Rule thresholds are set on in-domain normative data only, so the pathological and out-of-domain cohorts are never used for tuning.
Baselines
The naive reference point is the normative SynthSeg model applied to everything. It defines both the neurotypical ceiling and the pathological floor. Shang et al. report a mean Dice of 0.60 for the baseline on the pathological domain.
The strongest existing method is the pathology-weighted SynthSeg of Shang et al., which reports 0.67 mean Dice on the 33-subject pathological domain. It is the strongest competitor because it is the only method in the file evaluated on exactly this combined pathology-plus-protocol shift. We found no reported number on Lacuna for the size of its neurotypical drop, so we will measure that ourselves.
Both models are run as released, or retrained with the authors' recipe on identical synthetic data, compute and tuning budget. All systems are evaluated on the same in-domain (FeTA), similar-domain (NesVor), out-of-domain (external hospitals, down to 0.55T) and pathological splits. These are subject to data access; if the pathological cohort is unavailable, we use the pathological cases within FeTA and state the substitution.
Beating the strongest method means two things together. Pathological Dice must be at least 0.67, and neurotypical Dice must be statistically indistinguishable from the normative model, judged by bootstrap confidence intervals.
Three further references bound the routing question. A fixed average of the two models' probability maps tests whether routing matters at all. An oracle router that uses the true diagnosis label gives the upper bound of case-level routing. A simple pathology classifier trained on in-domain images gives a realistic case-level router.
The key ablations isolate the rules themselves. A learned router, a gradient-boosted classifier that predicts which model has higher Dice, receives exactly the same measurement features. If it matches the rule router under held-out pathology, the rules contribute only readability. We also remove the label–intensity rule and, separately, the acquisition rules, to show which family carries the gains.
For failure detection, a model ensemble's voxel disagreement and a learned Dice-regression quality-control model serve as baselines. No reported result was found on Lacuna for failure-detection performance in fetal MRI.
Why not just…?
Why not just train one model on a better mix of normal and pathological synthetic data? That is what Shang et al. did, and the trade-off it produced is the motivation for this work. The forgetting results of Mamalakis et al. suggest that one set of weights struggles to hold both populations. The averaged-ensemble and single-mixed-model baselines test this directly. If a carefully balanced single model reaches 0.67 on pathological brains with no neurotypical drop, routing is unnecessary, and we would report that.
Why not just route on network uncertainty from an ensemble? The failure we target is a model that is confidently normal on an abnormal brain. Pulling outputs back onto the healthy manifold is exactly the behavior that should produce agreement among ensemble members, not disagreement. This is an empirical claim. The failure-detection experiment compares rule verdicts with ensemble disagreement using AUROC for predicting per-structure Dice below 0.5, and if uncertainty wins, the claim fails.
Why not just learn the router or quality predictor from the same features? This is the most serious threat. On the training distribution, a learned router will likely match or beat the rules. Our bet is that the rules transfer better because they encode invariants (enclosure, compartment sums, self-relative intensity) rather than fitting the statistics of one pathology. The decisive test trains the learned router with ventriculomegaly cases available and evaluates it on spina bifida and on 0.55T scans. If the learned router still matches the rules there, the contribution narrows to interpretability. We would then present it that way, with the rules recast as a readable surrogate of the learned router.
Why not just route by diagnosis with an image classifier? Case-level routing cannot handle a brain that is abnormal in one structure and normal elsewhere, and it gives no account of why a given output is trusted. The oracle-diagnosis router bounds what case-level routing can achieve. Per-structure rule routing has to exceed that bound to justify itself.
Proposed Experiments
The main experiment reports Dice per structure and per domain for every system. It also gives the neurotypical penalty relative to the normative model, and coverage–risk curves showing the error remaining as abstention increases.
The first decisive check is the held-out-pathology comparison between the rule router and the learned router described above. The second is explanation validity. On roughly 60 cases, two raters, blinded to system identity, label the cause of each segmentation failure, and we measure agreement between the fired rule chain and the rated cause. For comparison, the raters also judge Grad-CAM-style maps of the segmenter on the same cases.
A smaller check reuses SABRE through the physics-informed setting. It asks whether the acquisition rules flag as unverifiable the SPGR gray-matter measurements that Gordaliza et al. found less consistent, without penalizing MPRAGE.
The project is worth continuing only under two conditions. The router must keep the 0.67 pathological gain with no significant neurotypical loss. And the rules must outperform the learned router on held-out pathology, or at least match it while producing explanations that raters judge correct more often than the saliency maps.
References
- Towards contrast- and pathology-agnostic clinical fetal brain MRI segmentation using SynthSeg
- Intensity scaling of conventional brain magnetic resonance images avoiding cerebral reference regions: A systematic review
- XAI-MeD: Explainable Knowledge Guided Neuro-Symbolic Framework for Domain Generalization and Rare Class Detection in Medical Imaging
- NEURO-GUARD: Neuro-Symbolic Generalization and Unbiased Adaptive Routing for Diagnostics - Explainable Medical AI
- Single Domain Generalization in Diabetic Retinopathy: A Neuro-Symbolic Learning Approach
- Neuro-Symbolic Spatial Reasoning in Segmentation
- Coherent Concept-based Explanations in Medical Image and Its Application to Skin Lesion Diagnosis
- Socrates: a flexible toolkit for building logic-based expert systems
- Healthcare Decision Support System for Administration of Chronic Diseases
- Physics-Informed Brain MRI Segmentation
- SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining
- A continually expandable foundation model for brain MRI