BrainIAC: Adaptive Brain MRI AI Systems

• BrainIAC is primarily a 3D framework for interactive brain-lesion segmentation that combines heterogeneous MRI inputs, clinician prompts, and online learning from corrected predictions.
• The segmentation system supports missing modalities through zero-filling and modality-agnostic channels, while Mid-Interaction and Post-Interaction adaptation improve accuracy and reduce required clicks under distribution shift.
• BrainIAC also names a separate 3D brain-MRI foundation-model encoder and an earlier Boolean associative-brain proposal, so researchers must distinguish these systems by context, architecture, and evidence base.

BrainIAC is a name used for more than one brain-related artificial-intelligence concept. In its current technical usage, it denotes Brain lesion Interactive Adaptive Continuously learning segmentation, a unified 3D framework for interactive brain-lesion segmentation, heterogeneous MRI modalities, and online model adaptation (Xu et al., 19 Sep 2026). The same name also appears as a comparator for a separate 3D brain-MRI foundation model in neuroimaging studies, while an earlier proposal used “BrainIAC” for an associative, memory-centered artificial-brain architecture based on Boolean circuits, state machines, and short-term memory (Burger, 2010). These usages are related only conceptually unless explicitly distinguished.

1. Namesake architectures and conceptual lineage

The principal contemporary BrainIAC framework is a 3D medical-image-segmentation system designed for deployment under distribution shift. Its name expands to Brain lesion Interactive Adaptive Continuously learning segmentation. The framework combines a multimodal 3D residual encoder U-Net, interactive prompts, heterogeneous-modality handling, and online adaptation driven by clinician corrections (Xu et al., 19 Sep 2026).

BrainIAC addresses three deployment conditions simultaneously: MRI protocols may provide different modality sets, clinicians may need to correct an individual segmentation interactively, and new scanners, pathologies, or acquisition conditions may require adaptation after deployment. Its distinctive premise is that interaction is not only an inference-time correction mechanism. Corrected predictions are also used as pseudo-labels to update the model for subsequent cases.

The name also occurs in the literature as a 3D brain-MRI foundation-model encoder. In an IDH-mutational-status benchmark, BrainIAC is described as an approximately 88.3-million-parameter 3D ViT-B/16 model pretrained on approximately 32,000 brain-MRI studies using SimCLR-style contrastive learning. It processes complete skull-stripped FLAIR and post-contrast T1-weighted volumes, producing modality-specific embeddings that are concatenated into a 1,536-dimensional patient-level representation (Hollet et al., 22 Jun 2026). Other studies evaluate this encoder for African-cohort generalization, MRI-artifact robustness, medical perceptual losses, and downstream representation transfer (Mouheb et al., 30 Jul 2026, Mielcarz et al., 6 Aug 2026, Fang et al., 28 Aug 2026, Akinmuleya et al., 21 Sep 2026).

A separate 2010 proposal, “Artificial Brain Based on Credible Neural Circuits in a Human Brain,” presents BrainIAC-like systems as associative processors rather than conventional von Neumann computers or weighted artificial neural networks (Burger, 2010). That architecture separates short-term memory, associative long-term memory, learned state machines, and nanoprocessing. It is a conceptual hardware proposal, not the medical-segmentation framework or the MRI foundation-model encoder.

2. Brain lesion Interactive Adaptive Continuously learning segmentation

BrainIAC constructs a common modality-channel space from the union of modalities present in its training databases. If the databases have modality sets $\mathbb{M}_1,\ldots,\mathbb{M}_N$, the number of input channels is

$C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$

Unavailable modalities are zero-filled into their corresponding channels. In the five-database configuration, the union is

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$

so $C=6$. The evaluated datasets include BraTS, ATLAS, MSSEG, WMH, and TBI, with modality sets ranging from T1-only data to multiparametric combinations containing FLAIR, T1, T1c, T2, PD, and SWI (Xu et al., 19 Sep 2026).

Zero-filling is combined with random modality dropping. If a sample contains $C_i$ available modalities, BrainIAC samples

$n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$

and replaces $n$ available modalities with zeros. This exposes the model to multiple modality subsets and discourages it from associating a fixed modality pattern with a particular pathology or database. The authors refer to this heterogeneous-modality strategy as Multi-Unet.

BrainIAC also introduces a modality-agnostic channel for sequences entirely absent during training. In the TBI experiments, SWI can be assigned to this channel when SWI was not included in the training configuration. The channel is trained using synthetic modality variations and information derived from other sequences. A randomly initialized extra channel produced little improvement, whereas the pretrained modality-agnostic channel allowed online adaptation to exploit SWI.

The segmentation backbone is a 3D residual encoder U-Net following the nnU-Net design. Its input has shape

$(C+3)\times H\times W\times D,$

where the additional three channels encode a bounding box, foreground clicks, and background clicks. The network produces a single-channel volumetric lesion mask. It supports four operational modes:

• automatic segmentation without prompts;
• bounding-box initialization;
• click-only interaction;
• bounding-box initialization followed by corrective clicks.

All scans are resampled to $1\times1\times1$ mm, standardized to $192^3$ voxels, and Z-score normalized. Full volumes are used.

3. Prompted inference and training objectives

A bounding-box prompt is represented as a binary volume. Although the clinician supplies a 2D box, BrainIAC extends it through three adjacent depth slices, creating a prompt region of size $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$0. Foreground and background clicks are encoded as one-valued voxels in separate channels, and clicks accumulate over interaction steps.

During training, bounding boxes are included with probability $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$1 and clicks independently with probability $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$2. The resulting prompt configurations are approximately:

Prompt configuration	Probability
Bounding box and clicks	49%
Bounding box only	21%
Clicks only	21%
No prompt	9%

Clicks are simulated as corrections. The network first predicts from the available initialization, after which false-negative and false-positive regions are identified. Clicks are sampled from false negatives, false positives, or a mixture of both, with equal probability across these strategies. The number of clicks is sampled uniformly from 1 to 10. Bounding boxes can include up to six connected components and receive 0–10% random padding.

The training objective is

$C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$3

where

$C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$4

The second term is the Click-Centered Gaussian loss (CCG). For a click at voxel $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$5, the spatial weighting function is

$C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$6

The Gaussian is multiplied by a class-limited indicator so that a foreground click emphasizes nearby foreground voxels and a background click emphasizes nearby background voxels. The reported settings are $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$7 and $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$8.

The purpose of CCG is to make a prompt influence a local region rather than only the clicked voxel. During ordinary training, the target mask is the ground truth. During adaptation, it is a pseudo-label derived from the model’s corrected prediction.

4. Online adaptation

BrainIAC maintains model parameters across sequential cases. If images arrive as $C=\left|\bigcup_{i=1}^{N}\mathbb{M}_i\right|.$9, interactions on earlier cases update the parameters used on later cases. The framework has two adaptation mechanisms: Mid-Interaction adaptation (MI) and Post-Interaction adaptation (PI).

Mid-Interaction adaptation

Let $$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$0 be the prompts before a new click $$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$1, and let

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$2

With parameters $$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$3, the prediction before the click is

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$4

and the prediction after adding the click is

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$5

BrainIAC treats $$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$6 as a pseudo-target for $$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$7. The MI objective is

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$8

A single gradient update gives

$$\{\mathrm{FLAIR},\mathrm{T1},\mathrm{T1c},\mathrm{T2},\mathrm{PD},\mathrm{SWI}\},$$9

The current prediction is then recomputed using $C=6$0. MI therefore changes the response to the current case and begins adaptation to the current distribution immediately.

Post-Interaction adaptation

After the clinician finishes a case, the final corrected prediction $C=6$1 is used as pseudo-supervision. PI adaptation has two stages.

First, the model generates an initial prediction without correction clicks, using no prompt or only the initial bounding box. A gradient update minimizes the Dice-plus-cross-entropy loss between that prediction and $C=6$2.

Second, a new click set is generated by comparing the updated initial prediction with $C=6$3. One click is placed in each false-positive or false-negative connected component, up to a specified maximum. The resulting click-conditioned prediction is optimized using the DCE and CCG terms. The updated parameters are retained for subsequent images.

MI and PI have different functions:

• MI provides immediate per-click refinement and adaptation during the current case.
• PI consolidates the completed case and improves initialization and click-conditioned correction on future cases.

On TBI FLAIR-visible lesions, the reported BBox+6 Dice scores were 47.9% without either mechanism, 61.6% with MI only, 59.1% with PI only, and 62.2% with both mechanisms.

5. Datasets, benchmarks, and empirical performance

BrainIAC was evaluated across seven datasets involving tumors, stroke, multiple sclerosis, white-matter hyperintensities, traumatic brain injury, vestibular schwannoma, and ischemic stroke.

Dataset	Pathology	Modalities	Cases
BraTS 2016	Brain tumor	FLAIR, T1, T1c, T2	444 train / 40 test
ATLAS	Stroke	T1	459 / 195
MSSEG	Multiple sclerosis	FLAIR, T1, T1c, T2, PD	37 / 16
WMH	White-matter hyperintensities	FLAIR, T1	42 / 18
TBI	Traumatic brain injury	FLAIR, T1, T2, SWI	156 / 125
VES-SEG	Vestibular schwannoma	T1c, T2	242 test-only
ISLES	Ischemic stroke	FLAIR, T1, T2, DWI	28 test-only

Three training configurations were used: a four-database model holding out MSSEG, a four-database model holding out TBI, and a five-database model evaluated on VES-SEG and ISLES. Held-out datasets were treated as out-of-distribution evaluation and online-adaptation data.

On held-out MSSEG using FLAIR, BrainIAC without online adaptation achieved 53.6% Dice with no prompt, 57.3% with a box, and 68.5% with BBox+6. Online adaptation increased these values to 64.7%, 66.2%, and 72.8%, respectively. With FLAIR and T1, the no-prompt score increased from 55.0% to 66.1%, while BBox+6 increased from 69.6% to 73.3%.

On held-out TBI using FLAIR, the reported scores were:

Prompt setting	Without adaptation	With adaptation
No Prompt	37.9%	44.3%
BBox	39.5%	50.5%
BBox+1	41.1%	54.3%
BBox+6	47.9%	62.2%

With FLAIR, T1, and T2, BBox+6 increased from 52.4% without adaptation to 62.7% with adaptation. For all lesions, including SWI-visible lesions, the adapted FLAIR-only model reached 58.0% at BBox+6, while the adapted FLAIR/T1/T2 configuration reached 58.7%.

For the unseen SWI modality, assigning SWI to the modality-agnostic channel produced 61.0% Dice at BBox+6, compared with 58.6% when the channel was zero-filled. An alternate 12-click strategy detected 68.9% of SWI-only components, compared with 53.1% without SWI access, with an overall Dice of 61.2%.

The largest adaptation effect was reported on VES-SEG. Interactive segmentation improved from 6.5% automatically to 51.9% after BBox+6 without online adaptation, and to 83.8% with online adaptation. On ISLES, the reported BBox+6 improvement was approximately 65.2% to 70.9%.

In the five-database in-distribution setting, the reported No Prompt/BBox+6 Dice values without online adaptation were:

Dataset	No Prompt	BBox+6
MSSEG	74.4%	77.7%
TBI	59.9%	66.9%
WMH	78.0%	79.6%
ATLAS	58.2%	75.1%
BraTS	92.5%	93.2%

The framework was compared with nnInteractive, MedSAM2, IA+SA, and TSCA. IA+SA and TSCA used the same BrainIAC backbone, allowing the comparison to emphasize adaptation strategy. BrainIAC’s online version exceeded the reported IA+SA and TSCA results in the relevant held-out settings.

6. Interaction efficiency, adaptation behavior, and deployment

Online adaptation reduced the number of clicks needed to reach pathology-specific Dice thresholds.

Dataset	Average clicks without adaptation	With adaptation	Threshold reached without	With adaptation
MSSEG	6.66	3.81	71.7%	96.2%
VES-SEG	12.81	2.67	31.8%	90.1%
ISLES	5.46	3.43	64.3%	85.7%
TBI	10.35	6.78	39.9%	69.8%

The thresholds were 0.70 for MSSEG, 0.80 for VES-SEG, and 0.65 for ISLES and TBI. In an MSSEG component-recovery experiment with unlimited clicks, the adapted model required an average of 3.15 clicks to recover all missed components, compared with 20.94 without adaptation.

The implementation processes full $C=6$4 volumes with batch size one. Online MI and PI updates use Adam with learning rate $C=6$5 and one gradient step per update. On VES-SEG, MI adaptation required approximately 1.19 seconds per click and PI adaptation approximately 3.54 seconds per case. PI-only adaptation can be used when immediate per-click updates are undesirable.

A 3D Slicer plug-in is intended to provide the deployment interface. It supports volume visualization, bounding boxes, foreground and background clicks, evolving 3D masks, and sequential online adaptation. The framework is therefore designed as an interactive clinical workflow rather than solely as an offline segmentation model.

The comparisons with nnInteractive and MedSAM2 require contextual qualification. nnInteractive is a large 3D interactive model trained on more than 120 datasets but processes one modality at a time and does not perform online adaptation. MedSAM2 is a 3D SAM-based model supporting box prompts, also with single-modality processing and frozen inference. MedSAM2’s evaluation used a tumor bounding box from the middle slice as a prompt, which may provide an advantage relative to fully automatic deployment.

7. BrainIAC as an MRI foundation-model encoder

The foundation-model usage of BrainIAC is distinct from the interactive segmentation framework. In the IDH benchmark, BrainIAC processes full skull-stripped FLAIR and T1c volumes at 1-mm isotropic resolution. The two modality embeddings are concatenated and supplied to a class-weighted linear classifier. BrainIAC is evaluated as a frozen feature extractor rather than fine-tuned end to end (Hollet et al., 22 Jun 2026).

Across four adult glioma cohorts, BrainIAC consistently underperformed the strongest radiomics and broader image-foundation-model baselines. TabPFN applied to radiomics achieved mean AUROC $C=6$6 and AUPRC $C=6$7, while BiomedCLIP achieved approximately $C=6$8 AUROC and BrainDINO approximately $C=6$9. The paper does not provide a pooled numerical AUROC or AUPRC for BrainIAC. Its calibration was poor: BrainIAC’s ECE was $C_i$0 on UCSF-PDGM, $C_i$1 on UPENN-GBM, $C_i$2 on EGD, $C_i$3 on UTSW-Glioma, $C_i$4 cross-cohort, and $C_i$5 on the external UCSD-PTGBM cohort.

Other evaluations similarly show that BrainIAC’s performance is task- and distribution-dependent. On a Nigerian dementia-versus-control task, BrainIAC achieved ROC-AUC $C_i$6 with linear probing and $C_i$7 with LoRA; on OASIS-4 it achieved $C_i$8 and $C_i$9, respectively (Mouheb et al., 30 Jul 2026). In a three-way Nigerian classification task involving Control, Dementia, and Parkinson’s disease, its frozen representation produced accuracy values between 0.351 and 0.400 across T1w, T2w, T1w+T2w, and FLAIR, with MCC values from $n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$0 to $n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$1 (Akinmuleya et al., 21 Sep 2026). For T1w+T2w, class-specific F1 values were 0.028 for Control, 0.545 for Dementia, and 0 for Parkinson’s disease.

BrainIAC is also sensitive to simulated MRI artifacts. In a controlled BraTS-Africa study involving k-space spikes, periodic ghosting, Gibbs ringing, Rician noise, Gaussian blur, bias fields, and gamma contrast, its CKA decreased substantially under many corruptions and approached zero for several frequency-domain perturbations (Mielcarz et al., 6 Aug 2026). RankMe remained comparatively stable, indicating geometric distortion without consistent dimensional collapse. Periodic ghosting and Rician noise strongly degraded both BrainIAC representations and an independent TumorSynth segmentation-consistency measure, although representation-level and task-level robustness were only partially aligned.

In a study of perceptual losses for brain-MRI contrast-dose simulation, BrainIAC was evaluated as a frozen candidate alongside RadImageNet, SegVol, VGG16, and ResNet50. It performed worst in the four-task aggregate linear-probing rank, with mean rank 4.37, whereas RadImageNet achieved 1.84 (Fang et al., 28 Aug 2026). RadImageNet, not BrainIAC, was therefore selected for the dose-simulation experiment. Replacing VGG16 with RadImageNet increased PSNR from 41.63 to 41.74 and SSIM from 0.9739 to 0.9754, but no BrainIAC perceptual-loss experiment was performed.

BrainDINO and BrainG3N provide further comparative context. BrainDINO is a 2D self-distilled brain-MRI model trained on approximately 6.6 million slices; under frozen transfer it outperformed BrainIAC on many tasks, including brain-age estimation, IDH prediction, and several segmentation benchmarks, although BrainIAC remained superior for some meningioma segmentation endpoints (Wu et al., 30 Apr 2026). BrainG3N is a volumetric masked-autoencoder tokenizer and outperformed BrainIAC or matched it on 21 of 23 reported probing tasks, including a brain-age MAE of 4.43 years versus 7.33 years for BrainIAC in the representative table (Puyvelde et al., 17 Jun 2026).

These results do not establish that the BrainIAC encoder is intrinsically unsuitable for all neuroimaging applications. They evaluate particular frozen-feature protocols and may reflect pretraining scale, objective, domain composition, input handling, probe capacity, and acquisition mismatch. Full fine-tuning, LoRA adaptation, site-specific calibration, richer multimodal heads, and locally representative pretraining may produce different results.

8. Earlier associative BrainIAC proposals

The 2010 artificial-brain proposal differs fundamentally from both contemporary BrainIAC usages. It presents an associative processor composed of:

• short-term memory identified with consciousness;
• associative long-term memory identified with subconscious memory;
• state machines embedded in long-term memory;
• nanoprocessing within short-term memory;
• sensory input, cue editing, cue gating, importance encoding, and control logic (Burger, 2010).

Its causal loop is

$n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$2

The proposal translates neurons into Boolean CMOS logic using a “substitution principle.” It distinguishes ordinary logical neurons, short-term-memory neurons, and long-term-memory neurons, with persistence ranging from tens of milliseconds to several seconds and indefinitely, respectively. Long-term memory is modeled as a PROM-like set-only latch, while short-term memory is modeled using charge storage and fast and slow currents.

Learned procedures are represented as state machines:

$n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$3

The paper illustrates equation solving through associative recall and state-machine manipulation of distributed Boolean attributes. For example,

$n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$4

is transformed to $n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$5 and then $n\sim \mathrm{Uniform}\{0,\ldots,C_i-1\}$6 through memory movement, associative subtraction, division lookup, and local reversible operations. It also proposes landmark navigation, left-right transformations, and nanocode controlled by “TO” and “FM” bits.

This architecture is not empirically validated as a completed BrainIAC implementation. It reports no fabricated prototype, benchmark, simulation results, power analysis, fabrication result, or learning algorithm. Its relevance to contemporary BrainIAC systems is conceptual: it emphasizes memory-centered computation, distributed representations, learned procedures, feedback, and physically grounded control rather than conventional software-only reasoning.

The medical-segmentation BrainIAC framework retains one abstract parallel with this earlier proposal: both treat adaptation and feedback as central to intelligent behavior. The segmentation system, however, is a neural image-analysis model with pseudo-label-based online learning, not a Boolean artificial brain.

9. Scope, limitations, and significance

The contemporary segmentation BrainIAC framework demonstrates a specific form of adaptive medical AI: 3D lesion segmentation that can accommodate heterogeneous modality sets, accept clinician prompts, and update itself online from interaction-derived pseudo-labels. Its strongest reported properties are the complementarity of modality handling, interactive correction, and adaptation under distribution shift. The largest empirical improvements occur on held-out pathologies, unseen modalities, and small or heterogeneous datasets.

Several limitations constrain interpretation. Interaction was simulated rather than conducted by radiologists. Online adaptation can propagate erroneous pseudo-labels if the final corrected prediction is inaccurate. The model performs binary lesion segmentation rather than distinguishing multiple lesion classes or tumor subregions. Unseen-modality use requires a modality-agnostic channel and interaction; it is not automatic. Full-volume inference imposes computational costs, and prospective multi-site clinical validation remains necessary.

The foundation-model BrainIAC is a separate object of evaluation. Its performance varies substantially by downstream task, pretraining distribution, artifact type, and adaptation protocol. It can be competitive in some settings, but frozen representations may collapse toward majority classes, exhibit poor calibration, or be sensitive to MRI artifacts and population shifts. Comparisons with BrainDINO, BrainG3N, BiomedCLIP, RadImageNet, radiomics, and end-to-end task-specific models show that domain-specific pretraining alone does not guarantee generalization.

The earlier associative architecture is more speculative still. It provides a hardware and cognitive-computation proposal but does not demonstrate human-like intelligence, scalable associative memory, biologically faithful neural abstraction, or a working artificial brain.

Accordingly, “BrainIAC” should be disambiguated by context. In current medical-image analysis, it most commonly denotes Brain lesion Interactive Adaptive Continuously learning segmentation, a multimodal interactive segmentation framework with MI and PI online adaptation (Xu et al., 19 Sep 2026). In neuroimaging representation learning, it denotes a separate 3D brain-MRI foundation-model encoder evaluated through frozen-feature transfer (Hollet et al., 22 Jun 2026). In historical artificial-intelligence discussions, it can refer to an associative circuit architecture based on Boolean logic, memory, state machines, and nanoprocessing (Burger, 2010). These systems share an emphasis on adaptive, brain-related computation, but they differ in task, architecture, evidence base, and technological maturity.