Research map / Detection and segmentation / Open-set and few-shot detection
Bi-level Learning of Task-Specific Decoders for Joint Registration and One-Shot Medical Image Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Bi-JROS, a bi-level optimization framework with a frozen pretrained shared encoder and task-specific registration/segmentation decoders, plus an appearance conformity constraint for one-shot brain MRI segmentation.
Problem
One-shot medical image segmentation via joint registration and segmentation suffers from unstable training due to naive joint or alternating optimization of task-specific decoders, and from appearance mismatch between labeled and unlabeled data used for augmentation.
Approach
Model registration decoder as the upper-level leader and segmentation decoder as the lower-level follower in a bi-level optimization with an implicit-function-based gradient response algorithm, using a frozen pretrained shared encoder and an appearance conformity constraint that warps the fake-labeled image via bidirectional deformations instead of the labeled image.
Key results
- Bi-JROS reaches 82.8% segmentation Dice and 80.8% registration Dice with 0.387 NCC on the mixed dataset, the best among compared JRS methods (BRBS: 81.8%, 80.0%, 0.324).
- On the new OASIS dataset, Bi-JROS achieves 81.4% segmentation Dice vs. 80.3% for BRBS while training in 7.8 hours vs. 9.6 hours for BRBS.
- Ablation shows bi-level/gradient response adds 0.6% registration and 0.7% segmentation Dice over alternating training; total gains over baseline are 1.7% (registration) and 1.6% (segmentation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Bi-JROS instead of BRBS | one-shot medical image segmentation · Mixed dataset | Segmentation Dice (%): 81.8 ± 2.4 -> 82.8 ± 2.2 | Helped |
| use Bi-JROS instead of BRBS | joint registration and segmentation · Mixed dataset | Registration Dice (%): 80.0 ± 2.4 -> 80.8 ± 2.0 | Helped |
| use Bi-JROS instead of SST | image registration · Mixed dataset | NCC: 0.364 ± 0.005 -> 0.387 ± 0.007 | Helped |
| use Bi-JROS instead of BRBS | one-shot medical image segmentation · OASIS | Segmentation Dice (%): 80.3 ± 1.1 -> 81.4 ± 0.6 | Helped |
| use Bi-JROS instead of BRBS | image registration · OASIS | Registration Dice (%): 79.5 ± 1.0 -> 80.0 ± 0.6 | Helped |
| use Bi-JROS instead of naive alternative training | image registration · OASIS | NCC: 0.356 ± 0.004 -> 0.362 ± 0.004 | Helped |
| freeze the encoder and alternately train the two decoders instead of joint training (ablation) | Bi-JROS · Mixed dataset | Registration Dice (%): 79.1 ± 2.0 -> 79.8 ± 2.0; 0.7% | Helped |
| add the appearance conformity constraint (ACC) (ablation) | Bi-JROS with a frozen encoder and alternating decoder training · Mixed dataset | Segmentation Dice (%): 81.5 ± 1.9 -> 82.1 ± 2.1; 0.6% | Helped |
| add bi-level optimization with Gradient Response (ablation) | Bi-JROS with ACC · Mixed dataset | Registration Dice (%): 80.2 ± 2.1 -> 80.8 ± 2.0; 0.6% | Helped |
| add bi-level optimization with Gradient Response (ablation) | Bi-JROS with ACC · Mixed dataset | Segmentation Dice (%): 82.1 ± 2.1 -> 82.8 ± 2.2; 0.7% | Helped |
Limitations
- Evaluated only on brain MRI datasets
- Gradient response relies on a Gauss-Newton style first-order approximation rather than exact bi-level gradients
- Improvements over the strongest baseline (BRBS) are modest, around 1 Dice point
Experimental setup
Datasets: ABIDE ADNI PPMI OASIS
Benchmarks: mixed brain MRI dataset (ABIDE, ADNI, PPMI) OASIS
Metrics: Dice NCC
Baselines: SyN NiftyReg deedsBCV VoxelMorph LKU-Net TransMorph UNet MASSL CPS SST DeepAtlas DataAug
Builds on: BRBS UReSNet DataAug VoxelMorph
Closest papers by meaning
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.81)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.81)
- Show and Segment: Universal Medical Image Segmentation via In-Context Learning (CVPR 2025 · similarity 0.81)
- Training Like a Medical Resident: Context-Prior Learning Toward Universal Medical Image Segmentation (CVPR 2024 · similarity 0.80)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.80)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.