DuMeta++: Meta-Learning for Brain Segmentation

• DuMeta++ is a spatiotemporal dual meta-learning framework that learns age-agnostic brain tissue features and enables rapid few-shot adaptation across different age domains.
• It employs a memory-bank-based class-aware regularization strategy to enforce longitudinal consistency and improve segmentation accuracy from pediatric to elderly MRI scans.
• It extends prior DuMeta work by incorporating prototype-based regularization and formal convergence proofs, yielding superior performance on cross-age brain tissue segmentation benchmarks.

Searching arXiv for the specified DuMeta++ paper and the earlier DuMeta paper to ground the article in current preprints. DuMeta++ is a spatiotemporal dual meta-learning framework for generalizable few-shot brain tissue segmentation across diverse ages, introduced to address age-related domain shift in MRI and the practical absence of paired longitudinal data (Sun et al., 6 Feb 2026). It combines meta-feature learning, which trains a universal encoder to produce age-agnostic anatomical features, with meta-initialization learning, which learns a favorable initialization for a lightweight segmentation head that can be rapidly fine-tuned in a new age domain while the encoder is frozen (Sun et al., 6 Feb 2026). The framework further incorporates a memory-bank-based class-aware regularization strategy to impose longitudinal consistency without explicit longitudinal subject pairs, and it is positioned as an extension of the earlier DuMeta formulation, which already coupled meta-feature learning and meta-initialization learning for one-shot brain tissue segmentation across the lifespan (Sun et al., 2023).

1. Problem setting and motivation

Brain tissue segmentation from MRI is affected by strong age-dependent appearance and morphological variation. In the formulation associated with DuMeta++, infants around six months exhibit isointense gray matter and white matter with near-overlapping intensities, whereas aging brains, including Alzheimer’s disease cases, show enlarged cerebrospinal fluid and cortical atrophy (Sun et al., 6 Feb 2026). These shifts degrade out-of-domain segmentation performance and also hinder longitudinal consistency.

The target setting is few-shot cross-age generalization. Clinical sites typically have only a few annotated scans in a target age domain, and the objective is to train on cross-sectional datasets from multiple ages and adapt with as few as one labeled volume to a new age domain while maintaining segmentation accuracy and longitudinal stability (Sun et al., 6 Feb 2026). In the earlier DuMeta formulation, this was expressed as one-shot brain tissue segmentation of CSF, GM, and WM across the human lifespan, with the additional challenge of preserving longitudinally consistent, age-agnostic representations after fine-tuning to a target age group (Sun et al., 2023).

A central motivation for DuMeta++ is the limited practicality of prior self-supervised approaches that depend on paired scans of the same subject across time. The method is explicitly designed to operate without paired longitudinal data and instead learns from cross-sectional datasets sampled across age-specific domains (Sun et al., 6 Feb 2026). This suggests a shift from explicit temporal pairing toward supervised or weakly supervised cross-domain anatomical alignment.

2. Core framework and architectural design

DuMeta++ is organized around a dual meta-learning framework with two coupled components (Sun et al., 6 Feb 2026). The first, meta-feature learning (MFL), treats the encoder as the meta-learner and optimizes it to be longitudinally consistent and robust across age domains while the segmentation head is adapted in an inner loop. The second, meta-initialization learning (MIL), learns a favorable initialization for a lightweight segmentation head that can be rapidly fine-tuned with few shots on a new age domain while the encoder remains frozen (Sun et al., 6 Feb 2026).

The backbone is a 3D U-Net with five down-sampling and five up-sampling blocks, InstanceNorm3d, ReLU, deep supervision at multiple decoder scales, and a standard encoder–decoder design (Sun et al., 6 Feb 2026). The encoder serves as the meta-learner $E_\theta(\cdot)$, and the decoder or segmentation head $D_\omega(\cdot)$ serves as the base-learner. Age is treated as a domain variable, and tasks are sampled across age-specific datasets; multi-scale features from the decoder are used to impose class-aware regularization across ages (Sun et al., 6 Feb 2026).

A distinctive element is the “tri-dataset” episode design. In each meta-training iteration, one dataset is sampled for an inner-loop update that mimics few-shot fine-tuning of the head, and two different datasets are sampled for the outer loops: one updates the encoder through MFL and class-aware regularization, and the other updates the head initialization through MIL (Sun et al., 6 Feb 2026). This episode design explicitly simulates cross-age shifts.

Relative to the earlier DuMeta method, DuMeta++ adds a memory-bank prototype design replacing sample-level pulls and pushes, introduces formal convergence proofs, and expands ablations and benchmarks (Sun et al., 6 Feb 2026). The earlier method also relied on dual meta-learning, but its class-aware regularization was formulated as two supervised multi-scale regularizers: inter-tissue spatial orthogonality and intra-tissue temporal similarity (Sun et al., 2023). The progression from DuMeta to DuMeta++ therefore centers on prototype-based regularization and theoretical analysis rather than a change in the basic dual meta-learning decomposition.

3. Mathematical formulation and optimization

The task distribution is defined over age domains $A$, including datasets such as IBIS-12M, IBIS-24M, and OASIS3, with tasks $\tau \sim T$ sampled across $A$ (Sun et al., 6 Feb 2026). For each task $\tau$, support $S_\tau$ and query $Q_\tau$ are sampled from the chosen dataset or datasets.

For an image $x$ with label $y$, prediction $D_\omega(\cdot)$0, and segmentation loss,

$D_\omega(\cdot)$1

In practice, the inner-loop and outer-loop segmentation objectives are Dice plus cross-entropy with deep supervision weights at multiple scales (Sun et al., 6 Feb 2026).

In MFL, the inner loop fine-tunes the head while the encoder is frozen: $D_\omega(\cdot)$2 where $D_\omega(\cdot)$3 is the inner-loop step size (Sun et al., 6 Feb 2026). The encoder is then updated through an outer-loop objective that accounts for the implicit dependence of the adapted head on the encoder: $D_\omega(\cdot)$4 with the derivative decomposed into a direct term and an implicit term through $D_\omega(\cdot)$5 (Sun et al., 6 Feb 2026). This bilevel structure is intended to train the encoder so that its representations remain effective after head adaptation under age shift.

In MIL, the encoder is fixed and the head is initialized from $D_\omega(\cdot)$6: $D_\omega(\cdot)$7 The outer-loop update is

$D_\omega(\cdot)$8

with

$D_\omega(\cdot)$9

In practice, DuMeta++ uses a first-order variant and drops the second-order term for stability and efficiency (Sun et al., 6 Feb 2026).

The earlier DuMeta paper presented an analogous bilevel decomposition and explicitly described the second-order term in the MIL step as omitted in practice to reduce computation and avoid instability, akin to first-order MIL variants such as Reptile (Sun et al., 2023). This continuity indicates that the computationally practical realization of the framework has consistently favored first-order meta-gradients for the head.

4. Memory-bank class-aware regularization

DuMeta++ introduces a memory-bank-based class-aware regularization strategy that uses per-class prototypes stored in a memory bank to enforce intra-class temporal alignment and inter-class spatial separation without requiring explicit longitudinal supervision (Sun et al., 6 Feb 2026). For $A$0 decoder scales, let $A$1 be the feature map and $A$2 the downsampled tissue label map. Per-class mean features $A$3 are computed for $A$4, and per-class prototypes $A$5 are maintained in a memory bank of capacity $A$6: $A$7

Using cosine distance

$A$8

DuMeta++ defines a margin-based triplet loss per tissue and scale. For GM on dataset $A$9 at scale $\tau \sim T$0,

$\tau \sim T$1

Analogous terms are defined for WM and CSF, and for dataset $\tau \sim T$2 (Sun et al., 6 Feb 2026). The full regularizer is aggregated across scales and tissues: $\tau \sim T$3 The encoder outer-loop loss becomes

$\tau \sim T$4

with $\tau \sim T$5 in the experiments (Sun et al., 6 Feb 2026).

This regularization replaces the earlier DuMeta pair of class-aware regularizers. In that earlier formulation, inter-tissue spatial orthogonality penalized cosine similarity between class prototypes at each scale, and intra-tissue temporal similarity encouraged cosine similarity of same-tissue prototypes across age-group datasets (Sun et al., 2023). DuMeta++ preserves the same conceptual aims—inter-class separation and cross-age same-class alignment—but re-expresses them through a memory-bank prototype mechanism with a margin-based triplet form (Sun et al., 6 Feb 2026).

The ablations reported for DuMeta++ indicate that adding MFL improves over baseline, adding MIL further improves, and prototype plus triplet memory-bank regularization yields the best results; removing the margin-based max or replacing prototypes with sample-level features degrades performance (Sun et al., 6 Feb 2026). A plausible implication is that the memory bank stabilizes cross-dataset prototype estimates and makes the longitudinal constraint less sensitive to mini-batch-level noise.

5. Training procedure, adaptation protocol, and datasets

During meta-training, DuMeta++ proceeds as follows (Sun et al., 6 Feb 2026). The encoder $\tau \sim T$6 and head $\tau \sim T$7 are initialized randomly. Each iteration performs: an inner-loop shared update of $\tau \sim T$8 on a sampled dataset using the segmentation loss; an encoder outer loop that samples mini-batches from two other datasets and updates $\tau \sim T$9 using $A$0; a head outer loop that, with encoder frozen, samples mini-batches from two datasets and updates the initialization $A$1 using first-order meta-gradients; and a memory-bank maintenance step that updates per-class mean features at each scale as running averages over bank capacity $A$2 (Sun et al., 6 Feb 2026).

At adaptation time on a new age domain, the encoder $A$3 is frozen, the head $A$4 is initialized from the learned $A$5, and only the head is fine-tuned with $A$6 shots, where $A$7 or $A$8, using the segmentation loss and deep supervision (Sun et al., 6 Feb 2026). The earlier DuMeta study further reported that fine-tuning only the last three up-sampling layers yields the best trade-off between convergence and accuracy, whereas freezing all layers is suboptimal and fine-tuning all layers improves performance but increases cost (Sun et al., 2023).

The experimental setup in DuMeta++ uses OASIS3, IBIS, ADNI, and iSeg-2019 (Sun et al., 6 Feb 2026). OASIS3 provides 1,639 T1-weighted MRIs from 992 individuals aged 42–95 and is used for meta-training. IBIS provides 1,272 T1w/T2w MRIs from 552 infants aged 3–36 months, with IBIS-12M and IBIS-24M used for meta-training. ADNI provides 2,389 longitudinal T1w MRIs aged 20–90 and is used for meta-testing. iSeg-2019 consists of 6-month infant T1w/T2w data with 10 training and 13 test subjects and is also used for meta-testing (Sun et al., 6 Feb 2026).

Preprocessing consists of skull stripping, bias field correction, and intensity normalization. Pseudo-labels for meta-training are generated with iBEAT, whereas meta-testing labels are refined by experts to form gold standards (Sun et al., 6 Feb 2026). Meta-training datasets are split 80/20 into train and validation. At meta-test, one training subject per target dataset is used to fine-tune the head, and evaluation is performed on the test sets under one-shot and five-shot protocols (Sun et al., 6 Feb 2026).

Implementation details include a 3D U-Net with InstanceNorm3d and ReLU, SGD with Nesterov momentum $A$9, polynomial learning-rate decay from $\tau$0, weight decay $\tau$1, input patches of size $\tau$2, batch size $\tau$3, deep supervision at five scales with weights $\tau$4, and nnU-Net-style augmentations (Sun et al., 6 Feb 2026). Evaluation uses Dice similarity coefficient and Average Symmetric Surface Distance, and longitudinal consistency is assessed by STCS and ASPC (Sun et al., 6 Feb 2026).

6. Theoretical properties and empirical performance

DuMeta++ provides formal convergence guarantees for both outer loops (Sun et al., 6 Feb 2026). For the encoder outer loop, assuming $\tau$5 is $\tau$6-Lipschitz smooth and $\tau$7 is $\tau$8-Lipschitz, with monotone step sizes

$\tau$9

satisfying $S_\tau$0, $S_\tau$1, $S_\tau$2, and $S_\tau$3, the method achieves

$S_\tau$4

and more specifically

$S_\tau$5

for a constant $S_\tau$6 independent of $S_\tau$7 (Sun et al., 6 Feb 2026). An analogous theorem holds for the head outer loop under corresponding conditions on $S_\tau$8 and $S_\tau$9 (Sun et al., 6 Feb 2026). The proof sketch treats the updates as stochastic gradient steps with unbiased noise due to mini-batch sampling and derives a non-asymptotic bound on the minimum expected gradient norm over $Q_\tau$0 iterations (Sun et al., 6 Feb 2026).

Empirically, under one-shot evaluation on iSeg-2019, DuMeta++ achieves Dice of CSF $Q_\tau$1, GM $Q_\tau$2, and WM $Q_\tau$3, with ASD of CSF $Q_\tau$4, GM $Q_\tau$5, and WM $Q_\tau$6 (Sun et al., 6 Feb 2026). On one-shot ADNI, it reaches Dice of CSF $Q_\tau$7, GM $Q_\tau$8, and WM $Q_\tau$9, with ASD of CSF $x$0, GM $x$1, and WM $x$2 (Sun et al., 6 Feb 2026). Under five-shot evaluation on ADNI, the reported Dice scores are CSF $x$3, GM $x$4, and WM $x$5, with ASD values of CSF $x$6, GM $x$7, and WM $x$8 (Sun et al., 6 Feb 2026).

For longitudinal consistency on ADNI, DuMeta++ attains STCS of CSF $x$9, GM $y$0, and WM $y$1, with ASPC of CSF $y$2, GM $y$3, and WM $y$4 (Sun et al., 6 Feb 2026). The paper reports that these results improve temporal stability compared to SSL baselines and DuMeta (Sun et al., 6 Feb 2026).

A concise comparison with the earlier DuMeta results clarifies the reported progression.

Setting	DuMeta (Sun et al., 2023)	DuMeta++ (Sun et al., 6 Feb 2026)
iSeg-2019 one-shot WM Dice	0.9145 ± 0.0126	0.9196 ± 0.0130
ADNI one-shot GM Dice	0.9678 ± 0.0034	0.9731 ± 0.0036
ADNI CSF STCS	0.9373	0.9393

The paper further reports that DuMeta++ outperforms PatchNCE, LSRL, LSOR, MAML, and Reptile under cross-age generalization benchmarks (Sun et al., 6 Feb 2026). Because the details block does not enumerate all corresponding numeric baseline values for every setting in DuMeta++, the most concrete comparative statement that can be made is that the trend is consistently favorable across the listed methods.

7. Interpretation, limitations, and relation to prior work

The rationale offered for DuMeta++ is threefold (Sun et al., 6 Feb 2026). First, the encoder learns semantically grounded, age-invariant features through meta-feature learning with implicit gradient coupling to the head’s inner-loop adaptation, fostering compatibility with future domain shifts. Second, class-aware regularization aligns per-class prototypes across ages and scales, directly shaping the feature space toward longitudinal consistency and tissue separability. Third, the meta-initialized head supports rapid few-shot adaptation using only decoder parameters, which is computationally efficient and less prone to overfitting when labeled data are scarce (Sun et al., 6 Feb 2026).

The framework is compared with two major lines of prior work. Relative to self-supervised longitudinal methods such as LNE, LSOR, and LSRL, DuMeta++ removes the requirement for paired scans and aligns features directly for segmentation (Sun et al., 6 Feb 2026). Relative to generic meta-learning baselines such as MAML and Reptile, it augments few-shot adaptation with explicit feature-level longitudinal constraints (Sun et al., 6 Feb 2026). The earlier DuMeta paper had already made a similar distinction, emphasizing that its supervised class-aware regularizations avoided pitfalls associated with unsupervised contrastive methods and that its encoder–decoder role separation differed from meta-learning the entire network (Sun et al., 2023).

Several limitations are explicitly identified for DuMeta++ (Sun et al., 6 Feb 2026). Pseudo-label reliance during meta-training may inject label noise and cohort bias. Single-prototype-per-class modeling may not capture intra-class heterogeneity. Freezing the encoder during adaptation may be restrictive under extreme domain shifts such as unusual pathology or very early neonatal stages. Validation is limited to T1w MRI and three tissue classes, and extension to multi-contrast imaging and broader clinical endpoints remains to be shown (Sun et al., 6 Feb 2026). These points suggest that robustness under more heterogeneous acquisition settings and pathology distributions remains an open research direction.

For deployment, the reported guidance is to pretrain DuMeta++ on available cross-sectional datasets spanning relevant ages, with pseudo-labels acceptable; for a new clinical site or age domain, freeze the encoder and fine-tune the head with $y$5-shot labeled scans, typically $y$6 to $y$7, using Dice plus cross-entropy and deep supervision; and, if continued training is planned, use memory-bank prototypes with capacity around $y$8, which is described as a good accuracy–efficiency trade-off (Sun et al., 6 Feb 2026). The ablations indicate best accuracy with margin $y$9 and regularizer weight $D_\omega(\cdot)$00, yielding mean Dice $D_\omega(\cdot)$01 and mean ASD $D_\omega(\cdot)$02 on ADNI, while memory-bank capacity in the range $D_\omega(\cdot)$03 to $D_\omega(\cdot)$04 offers only marginal gains beyond $D_\omega(\cdot)$05 (Sun et al., 6 Feb 2026).

Taken together, DuMeta++ represents a specific synthesis of bilevel meta-learning, class-aware prototype regularization, and few-shot decoder adaptation for lifespan brain tissue segmentation. Its principal technical claim is not merely improved benchmark performance, but that longitudinally consistent and age-agnostic segmentation can be achieved from cross-sectional training data alone, without paired longitudinal supervision, while retaining convergence guarantees and efficient adaptation behavior (Sun et al., 6 Feb 2026).