AI Research Portal

This paper proposes CausalBridgeNet, a causality-guided framework for adapting vision foundation models (VFMs) to medical image segmentation. The method introduces a Predictive Causal Reasoning Unit (PCRU) that estimates structured error maps and uses them to refine predictions iteratively without modifying the VFM backbone. The paper claims that CausalBridgeNet achieves state-of-the-art performance on the BraTS 2025 benchmark, demonstrating improved accuracy and robustness for high-precision medical segmentation.

The paper addresses several bottlenecks in applying vision foundation models (VFMs) to medical image segmentation:

Domain Shift: VFMs trained on natural images often perform poorly on medical images due to differences in modality, structure, and annotation style, leading to poor inductive bias and unreliable performance.
Computational Cost and Forgetting: Fine-tuning VFMs for medical tasks is computationally expensive and can lead to catastrophic forgetting of general visual priors.

Development of CausalBridgeNet, an efficient and scalable framework that enables VFMs to generalize effectively to 3D medical segmentation tasks without retraining or modifying the VFM backbone.
Design of the Predictive Causal Reasoning Unit (PCRU), which implements a closed-loop causal correction mechanism by estimating structured, class-specific prediction errors and refining outputs.
Demonstration of CausalBridgeNet's state-of-the-art performance on 3D brain tumor and 2D breast tumor segmentation benchmarks, highlighting its practical utility and clinical viability.

CausalBridgeNet is a plug-and-play framework designed to improve the cross-domain generalization of frozen VFMs for medical image segmentation. The process involves:

Vision Foundation Encoder: A hierarchical encoder is constructed by integrating frozen VFM blocks into a U-shaped architecture. This encoder progressively downsamples the input volume, preserving semantic richness across multiple scales. Lightweight modules like Gated Spatial Convolution (GSC), layer normalization, and MLP-based channel adapters are incorporated to support volumetric processing.
Semantic Decoder: This component reconstructs dense segmentation predictions by progressively upsampling the multi-scale encoder features using a U-shaped architecture. Skip connections fuse upsampled features with corresponding encoder outputs to recover fine-grained spatial information.
Predictive Causal Reasoning Unit (PCRU): This unit refines the initial predictions from the VFM backbone. It comprises two components:
- Mixture of Causal Experts (MoCE): This module estimates structured false positive (FP) and false negative (FN) maps. It takes intermediate decoder features and coarse segmentation logits, concatenates them with a learned prompt, and uses a router to select experts. The final output is a weighted sum of expert predictions, representing error signals.
- Causal Corrector: This component integrates the estimated FP and FN signals into the coarse segmentation logits. It applies additive adjustments, modulated by learnable scalar weights, to refine the predictions in a class-aware and direction-sensitive manner. This process mimics predictive error minimization.

The entire process operates with the VFM backbone frozen, making the approach efficient and scalable.

Architecture:
- The framework uses a U-shaped architecture incorporating a frozen VFM encoder and a semantic decoder.
- The encoder integrates Gated Spatial Convolution (GSC), grouped transformer blocks from the frozen VFM, layer normalization, and MLP-based channel adapters.
- The decoder uses 3D transposed convolutions and convolutional residual blocks.
- The PCRU consists of a Mixture of Causal Experts (MoCE) and a Causal Corrector.
Key Equations:
- Encoder feature computation:
  - El = GSC(El-1)
  - Ẽl = VFB(LN(Êl))
  - El = MLP(LN(Êl))
- Decoder feature computation:
  - Ďl = UpConv(Ďl+1)
  - Ďl = DecoderBlockl([Ďl, El])
- Raw segmentation logits: Y = Convout(D1)
- MoCE input: F = Concat(F, P)
- MoCE weights: w = Softmax(FC(GAP(F)))
- MoCE output: Scausal = ∑wk. Ek (F)
- Error maps: Scausal = [Sfp || Sfn]
- Causal Corrector refinement: Ŷc = Yc + βc. Sfc. Sfp,c
Loss/Objective:
- For BraTS2025: Dice loss and cross-entropy loss.
- For BUSI: Weighted sum of binary cross-entropy (BCE) and Dice loss.
Optimization:
- BraTS2025: Stochastic Gradient Descent (SGD) optimizer with a polynomial learning rate scheduler (initial learning rate 0.01, decay 1 × 10-5). Trained for 1,000 epochs.
- BUSI: Adam optimizer with cosine annealing (initial learning rate 1 × 10-4, minimum 1 × 10-5). Trained for 400 epochs.
Augmentation:
- BraTS2025: Brightness, gamma, rotation, scaling, mirror, and elastic deformation.
- General: Random rotation and flipping.
Data Handling:
- BraTS2025: 3D crop size of (64 × 64 × 64), batch size of 2.
- BUSI: Batch size of 8.
Implementation Details:
- Backbones used: CLIP and SAM2.
- The backbone model remains entirely frozen.
- Results are averaged over three random seeds for BraTS2025.

Datasets:
- BraTS2025 (Pre-only and Pre+Post configurations)
- BUSI (Breast Ultrasound Images)
Metrics:
- BraTS2025: Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95).
- BUSI: Intersection over Union (IoU) and F1 score.
Baselines:
- UNETR (Hatamizadeh et al. 2022)
- SwinUNETR (Hatamizadeh et al. 2021)
- SegResNet (Myronenko 2018)
- SegMamba (Xing et al. 2024)
- CausalBridgeNet (CLIP)
- U-Net, Att-Unet, U-Net++, U-NeXt, Rolling-UNet, U-Mamba, U-KAN (for BUSI)
Numeric Results:
- BraTS2025 Pre+Post: CausalBridgeNet (SAM2) achieved the best performance across all tumor subregions, outperforming SegMamba by +4.52 points in TC DSC and +5.96 points in RC DSC. It also showed significant improvements in HD95 scores.
- BraTS2025 Pre-only: CausalBridgeNet (SAM2) achieved the highest average DSC score of 91.16 and the lowest average HD95 of 3.58, outperforming all baselines. It demonstrated superior performance in segmenting the enhancing tumor (ET) with an HD95 of 3.82.
- Online Validation (BraTS2025): CausalBridgeNet (SAM2) achieved the highest DSC scores on ET (78.26), TC (80.05), and WT (91.96).
- BUSI: When integrated into U-KAN, the proposed correction module significantly improved IoU from 63.38% to 68.85% and F1-score from 76.40% to 80.86%.
Runtime/Training Cost: Not stated in the paper, but the method is described as efficient and scalable due to the frozen backbone.

No explicit paper-reported limitations were found.

Method evidence: "Inspired by the predictive coding theory of the Bayesian brain (Knill and Pouget 2004), which postulates that perception is driven by minimizing structured prediction errors, we propose to view domain adaptation through the lens of causal inference. In this view, the discrepancy between model prediction and input evidence can be interpreted as a structured error signal to guide refinement. Rather than relying on gradient-based fine-tuning of the backbone, we introduce a lightweight, interpretable correction mechanism that explicitly estimates and compensates for systematic prediction errors." (Section: Abstract)
Result evidence: "CausalBridgeNet achieves the best performance across all tumor structures, consistently outperforming previous models, including UNETER (Hatamizadeh et al. 2022), SwinUNETR (Hatamizadeh et al. 2021), SegResNet (Myronenko 2018), and SegMamba (Xing et al. 2024). Specifically, CausalBridgeNet (SAM2) achieves a DSC of 82.72 on TC and 82.12 on RC, surpassing SegMamba by +4.52 and +5.96 points, respectively. The HD95 scores also show significant improvement, indicating better boundary accuracy." (Section: Results on Pre+Post dataset)
Limitation evidence: Not stated in the paper.

Make Foundation Models Trustworthy Again: Causal Fine-Adaptation for Medical Image Segmentation

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures