Organizations: Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong, China · PET/CT Center, The First Affiliated Hospital of Xi’an Jiaotong University, No. 277 Yanta West Road, Xi’an 710061, China
Alzheimer's disease (AD) diagnosis may benefit from gut metabolic information associated with brain-gut interactions, yet existing models largely analyze cerebral glucose metabolism alone. Although whole-body fluorodeoxyglucose positron emission tomography (FDG-PET) can provide paired brain-gut observations, routine brain examinations rarely include gut imaging, limiting direct use of this information for diagnosis. In this paper, we present BRIDGE (Brain and Gut Representation Integration via Distribution-Guided Enhancement), a two-stage framework for transferring brain-gut metabolic knowledge to brain-only diagnosis. Specifically, BRIDGE integrates a conditional brain-to-gut latent transport module and a disease-evidence-guided cerebral pathway to capture subject-specific cross-organ correspondence and disease-relevant cerebral representations. For the cerebral pathway, a large language model (LLM) helps organize literature-derived descriptors into human-approved semantic anchors that guide feature learning. Additionally, we introduce a counterfactual-guided residual fusion mechanism, which uses matched and non-matched virtual-gut evidence to supervise a confidence gate and dynamically regulate the contribution of transported cross-organ information. At inference, BRIDGE exploits complementary gut-related context using brain PET alone, without measured gut PET. Extensive experiments across public and private cohorts demonstrate that BRIDGE outperforms brain-only and cross-organ baselines, achieving AUC gains of 6.0% and 5.2% over its brain-only counterpart for CN/MCI and CN/AD diagnosis, respectively. Notably, its ability to transfer complementary brain-gut metabolic context while preserving brain PET as the sole inference input highlights its potential for clinically practical diagnosis. Our code is available at https://github.com/CUHK-AIM-Group/BRIDGE.
Figures & tables
Fig. 1: Conceptual illustration and performance of BRIDGE. (a) Brain–gut representation learning and virtual-gut transfer for brain-only inference. (b) AUC comparison between the brain-only and +gut models for AD and MCI diagnosis on the public and private cohorts, respectively.
Fig. 2: Overview of BRIDGE framework. Stage 1 uses paired whole-body PET to learn subject-specific brain-to-gut latent transport through organ-specific encoders and endpoint constraints. Stage 2 uses labeled brain PET to train a cerebral pathway guided by literature-derived, human-approved semantic anchors and a virtual-gut residual pathway. An offline LLM assists in organizing the anchors, while matched and nonmatched gains supervise the confidence gate. At inference, a brain-only PET drives both pathways, and the gate regulates the virtual-gut contribution.
Cohort
Modality
Labels
N
Role
AutoPET [ 44 ]
BG
Unlabeled
188
Transport pretrain
PETWB [ 45 ]
BG
Unlabeled
484
Transport pretrain
QUADRA [ 46 ]
BG
Unlabeled
96
Transport pretrain
ENHANCE [ 47 ]
BG
Unlabeled
12
Transport pretrain
ADNI [ 48 ]
B
CN/MCI/AD
1,004
Source training validation
Private cohort
BG
CN/MCI
278
Primary evaluation
TABLE I: Cohort composition and experimental roles. BG and B denote paired brain–gut PET and brain PET, respectively. N denotes the number of examinations.
Method
Category
AUC ↑
ACC ↑
BACC ↑
SEN ↑
SPE ↑
Macro-F1 ↑
Brain-only baselines
ConvNeXt [ 18 ]
B/B
0.837 ± 0.050
0.728 ± 0.040 ∗
0.727 ± 0.039 ∗
0.816 ± 0.149
0.638 ± 0.101 ∗
0.726 ± 0.041 ∗
3D ViT [ 19 ]
B/B
0.684 ± 0.070 ∗
0.626 ± 0.068 ∗
0.626 ± 0.068 ∗
0.629 ± 0.090 ∗
0.624 ± 0.091 ∗
0.625 ± 0.069 ∗
Swin-T [ 20 ]
B/B
0.837 ± 0.060
0.756 ± 0.060
0.756 ± 0.060
0.743 ± 0.092
0.769 ± 0.071
0.755 ± 0.060
MiSePyNet [ 5 ]
B/B
0.736 ± 0.055 ∗
0.651 ± 0.075 ∗
0.653 ± 0.075 ∗
0.493 ± 0.203 ∗
0.812 ± 0.132
0.632 ± 0.095 ∗
Privileged-training comparators
TABLE II: Private-cohort CN/MCI classification performance (mean ± SD over five folds). B and BG denote brain and brain–gut PET. ∗ indicates significance by two-sided paired t -test with Holm correction ( p<0.05 ). Best and second-best results are shown in bold and underlined, respectively.
Method
Category
AUC ↑
ACC ↑
BACC ↑
SEN ↑
SPE ↑
Macro-F1 ↑
Brain-only baselines
ConvNeXt [ 18 ]
B/B
0.740 ± 0.039 ∗
0.661 ± 0.051 ∗
0.660 ± 0.049 ∗
0.631 ± 0.102 ∗
0.690 ± 0.165 ∗
0.656 ± 0.050 ∗
3D ViT [ 19 ]
B/B
0.691 ± 0.063 ∗
0.599 ± 0.061 ∗
0.596 ± 0.064 ∗
0.443 ± 0.267 ∗
0.750 ± 0.155 ∗
0.561 ± 0.130 ∗
Swin-T [ 20 ]
B/B
0.693 ± 0.072 ∗
0.619 ± 0.077 ∗
0.618 ± 0.080 ∗
0.551 ± 0.251 ∗
0.684 ± 0.168 ∗
0.601 ± 0.102 ∗
MiSePyNet [ 5 ]
B/B
0.935 ± 0.011
0.869 ± 0.021
0.869 ± 0.020
0.883 ± 0.069
0.855 ± 0.069
0.869 ± 0.021
Privileged-training comparators
TABLE III: ADNI CN/AD classification performance (mean ± SD over five folds). B and BG denote brain and brain–gut PET. ∗ indicates significance by two-sided paired t -test with Holm correction ( p<0.05 ). Best and second-best results are shown in bold and underlined, respectively.
Fig. 3: Private-cohort operating-point analysis. (a) Sensitivity and specificity across methods. Markers show five-fold means, error bars show SD, and marker size denotes AUC. The diagonal indicates equal sensitivity and specificity. (b) Expected error cost per 100 scans from five-fold mean sensitivity and specificity and cohort proportions, plotted against the missed-MCI penalty relative to false alarms. Red denotes BRIDGE, gray the comparators, and the navy dashed curve their lowest cost. Both axes are logarithmic, with lower error costs indicating better performance and the dotted line at 1 marking equal penalties.
Variant
AUC ↑
ACC ↑
BACC ↑
SEN ↑
SPE ↑
Macro-F1 ↑
Full BRIDGE
0.881 ± 0.050
0.818 ± 0.028
0.819 ± 0.027
0.821 ± 0.067
0.816 ± 0.029
0.815 ± 0.027
Component-level ablations
Brain-only
0.831 ± 0.030
0.766 ± 0.032
0.767 ± 0.032
0.764 ± 0.086
0.769 ± 0.069
0.766 ± 0.031
Without evidence anchors
0.875 ± 0.043
0.802 ± 0.028
0.802 ± 0.028
0.814 ± 0.073
0.790 ± 0.057
0.802 ± 0.027
Without latent transport
0.845 ± 0.054
0.774 ± 0.031
0.774 ± 0.031
0.757 ± 0.030
0.790 ± 0.044
0.774 ± 0.031
Fixed residual admission ( α=1 )
0.876 ± 0.027
0.791 ± 0.008
0.791 ± 0.008
0.800 ± 0.041
0.783 ± 0.037
0.791 ± 0.008
TABLE IV: Ablation study on private-cohort CN/MCI classification. Results are mean ± SD over five held-out folds. Component-level variants modify individual modules, whereas objective-level variants modify individual loss terms.
Fig. 4: Cross-organ transport alignment and task-specific feature geometry. (a) Cosine similarities of cerebral and transported virtual-gut latents to matched measured-gut latents ( n=278 ). The dashed line marks equality. (b) Paired gains after transport, with the point and bar indicating the mean and stratified-bootstrap 95% CI, respectively. (c) ADNI CN/AD fold-1 training features (182 CN, 174 AD). (d) Private-cohort CN/MCI fold-1 training features (96 CN, 98 MCI).
Fig. 5: Out-of-fold regional attribution on the private cohort. Circular bars show the ten AAL3 regions with the highest mean normalized positive integrated-gradient attribution to the final MCI logit, with scores rescaled to the top-ranked region. Surrounding surfaces localize the corresponding atlas parcels. Results aggregate all 278 examinations and reflect model-attributed diagnostic evidence rather than causal anatomy or imaging biomarkers.
Fig. 6: Illustrative brain–gut Grad-CAM maps. (a) ADNI AD brain PET, with attribution to the AD-class diagnostic logit. (b) Private-cohort MCI brain PET, with attribution to the MCI-class diagnostic logit. (c) Matched measured-gut PET from the same private examination, probed post hoc through the trained virtual-gut residual head. Columns show axial, sagittal, and coronal views.