AI Research Portal

This paper addresses the problem of confirmation bias in semi-supervised medical image segmentation, where initial errors in pseudo-labels can be reinforced and become difficult to correct. The authors propose DualFete, a dual-teacher feedback model that incorporates a feedback mechanism into the teacher-student paradigm. This mechanism allows the student to provide feedback on pseudo-labels, enabling the teachers to refine them and mitigate error propagation. The paper claims that DualFete effectively combats confirmation bias and improves segmentation performance on medical imaging benchmarks.

The paper identifies the following bottlenecks in semi-supervised medical image segmentation:

Error Propagation and Confirmation Bias: The teacher-student paradigm is vulnerable to erroneous pseudo-labels. The student's iterative reconfirmation of these errors leads to self-reinforcing bias, making it difficult to correct initial mistakes. This is particularly problematic in medical imaging due to inherent ambiguities.

Introduction of a feedback mechanism into the teacher-student framework to equip it with error-correction capabilities, thereby mitigating error propagation in semi-supervised medical image segmentation.
Development of a dual-teacher feedback model that integrates cross-supervision for conflict resolution and a feedback loss for preserving teacher diversity while suppressing errors, enhancing feedback dynamics.
Comprehensive experimental validation on three medical image benchmarks demonstrating the method's effectiveness in combating error propagation and confirmation bias.

The DualFete method enhances the teacher-student paradigm by introducing a feedback loop for error correction.

Feedback Mechanism: The core idea is that the student assesses whether updates driven by teacher-generated pseudo-labels align with ground-truth supervision on labeled data. This assessment is then fed back to the teacher to refine its pseudo-labels.
- Feedback Attributor: Identifies pseudo-labels responsible for triggering the student's update.
- Feedback Receiver: Determines which pseudo-labeling probabilities to adjust.
Feedback-Coupled Teacher-Student Model:
- The student's performance on a labeled mini-batch is evaluated.
- The change in student performance (δ) due to a pseudo-label update is calculated.
- A feedback loss (Lfь) is formulated to guide the teacher. If δ > 0 (performance improved), the teacher maximizes the likelihood of generating such pseudo-labels. If δ < 0 (performance worsened), the teacher reduces this likelihood. This mechanism aims to align pseudo-label updates with the supervised loss gradient.
Dual-Teacher Feedback Model: To enrich feedback dynamics and promote mutual learning, two teachers (φ and ψ) are employed.
- Pseudo-label Consensus: Pseudo-labels are set to the consensus if teachers agree, or to the higher-confidence label if they disagree.
- Attributors and Receivers: Feedback is quantified based on agreement (ŷa) and disagreement (ŷd) regions between teachers. Feedback is applied to either the lower-confidence side (for agreement) or the higher-confidence side (for disagreement) based on the feedback signal (δ).
- Dual-Teacher Feedback Loss (Ldf): A loss is formulated to modulate the pseudo-labeling likelihood of each teacher based on the feedback signals and receiver masks. This encourages productive disagreement and reduces consistent errors.
Holistic Framework:
- Student Training: The student is trained on unlabeled data using pseudo-labels generated by the dual-teacher model. It also provides feedback on labeled data.
- Dual-Teacher Training: Teachers are trained using a combination of fully-supervised loss on labeled data, the dual-teacher feedback loss (Ldf), and cross-supervised loss (LA) on unlabeled data.

Architecture: V-Net for LA and Pancreas datasets, U-Net for BraTS dataset.
Key Equations:
- Student loss: min Ls(θs,θT; Du) = min θs (1/Nu) Σu l(f(xu; θs), ŷu)
- Student performance evaluation: Li(θs; Di) = Σl(f(x1;θs), y1)
- Feedback signal: δ = Li(θs) – Li(θ's)
- Feedback loss: Lfb(θT; Du) = -Σ log P(ŷu|x²; θT, Du)
- Dual-teacher feedback loss: Ldf(θ) = -Σ δ log P(ŷu|x²; θ, Du, M)
- Teacher objective: min Ly(θ) = min θ Li(θ) + Laf(θ) + λLcs(θ; θ', A)
- Cross-supervised loss: LA(θ;θ', A) = (1/Nu) Σu l(f(x²; θ), A(ŷ'))
Loss/Objective:
- Student: Dice and cross-entropy combination for segmentation loss.
- Teachers: Fully-supervised loss (Li), dual-teacher feedback loss (Ldf), and cross-supervised loss (LA).
Optimization: SGD optimizer.
Augmentation: Strong augmentations (e.g., copy-paste, color jitter) are used.
Data Handling:
- Datasets: LA, Pancreas, BraTS19.
- Label settings: 5%, 10%, 20% for LA; 10%, 20% for Pancreas and BraTS.
- LA: 100 MRIs (80 train, 20 test).
- Pancreas: 82 CT scans (62 train, 20 test).
- BraTS: 335 brain MRIs (250 train, 25 validation, 60 test).
Implementation Details:
- For LA, normalized gradients are used in Eq. 7. For Pancreas and BraTS, this is not applied.
- A confidence threshold is employed to filter possibly unreliable targets in cross-supervised loss LA.
- A ramp-up weighted factor (λ) is used for the cross-supervised loss.

Datasets: LA, Pancreas, and BraTS19.
Metrics: Dice (%), 95% Hausdorff distance (95HD, voxel).
Baselines: Fully Supervise, UA-MT, SASSNet, DTC, URPC, PS-MT, MC-Net+, SS-Net, BCP, UniMatch, MutRel, AD-MT, TraCoCo.
Numeric Results:
- DualFete consistently outperforms existing approaches across nearly all benchmarks and label settings, even without fine-tuning.
- On Pancreas with 10% labels, DualFete achieves a +1.78% Dice improvement and 1.84-voxel 95HD reduction over previous best results (AD-MT).
- Ablation studies show that the feedback loss (Ldf) alone fails to maintain robustness against input perturbations and increases prediction uncertainty, while cross-supervision (LA) produces superior student performance.
- A confidence threshold of 0.7 for cross-supervised loss LA achieves optimal performance.
Runtime/Training Cost:
- Inference speed is comparable to other methods (approx. 1.9 s/case with TITAN RTX).
- Training time for DualFete is 2.28 s/iter, with 10.25 GB memory usage.

The paper speculates that oscillations in testing performance for the BraTS dataset might be due to the validation set being prone to overfitting due to its relatively limited quantity (25 samples).
Fine-tuning on very limited data is susceptible to overfitting, as observed on the LA dataset with 5% and 10% labels.

Method evidence: "In response, this work introduces a feedback mechanism into the teacher-student framework to counteract error reconfirmations. Here, the student provides feedback on the changes induced by the teacher's pseudo-labels, enabling the teacher to refine these labels accordingly." (Abstract)
Result evidence: "As shown in Table 1, DualFete outperforms existing approaches across nearly all benchmarks and label settings, even without fine-tuning. Notably, on Pancreas with 10% labels, it achieves a +1.78% Dice improvement and 1.84-voxel 95HD reduction over previous best results [AD-MT (Zhao et al. 2024)]." (Experimental settings)
Limitation evidence: "For the BraTS dataset, it is noteworthy that our DualFete shows some oscillations in terms of testing performance. We speculate that the validation set is prone to overfitting due to its relatively limited quantity (25 samples)." (Ablation studies)

DualFete: Revisiting Teacher-Student Interactions from a Feedback Perspective for Semi-supervised Medical Image Segmentation

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Evidence

Additional Figures