Research map / Label-efficient and robust learning / Domain adaptation and generalization
One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor 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 PSTUDA, a one-to-multiple unsupervised domain adaptation framework for kidney tumor segmentation across multi-sequence MRI, using a multi-level style dictionary and cascaded style fusion with point-wise instance normalization, and releases the MSKT dataset.
Problem
Annotation burden for kidney and tumor segmentation across multiple MRI sequences; existing UDA methods are one-to-one and inefficient for multi-target transfer.
Approach
A single shared generator translates a labeled source domain into multiple target MRI sequences using learnable per-domain multi-level style dictionaries and cascaded style fusion modules that recombine content and style per pixel via Point-wise Instance Normalization (PIN), with the generated images then used to train a U-Net segmenter.
Key results
- On MSKT, PSTUDA reaches average DSC 81.33% (T1c→FS T2W), 65.52% (T1c→T2W), and 80.61% (T1c→DWI), outperforming the compared UDA methods.
- PSTUDA reduces FLOPs by about 72% vs CycleGAN and 79% vs MUNIT, and needs about half the parameters (42.19M vs 79.80M for StarGAN v2 family), though FLOPs are 17% higher than StarGAN v2.
- On CT (KiTS19)→MRI, PSTUDA achieves average DSC 79.15 (T1c), 80.85 (FS T2W), 66.71 (T2W), 82.54 (DWI).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply PSTUDA rather than CycleGAN | Kidney tumor segmentation using T1c→FS T2W adaptation · MSKT: T1c→FS T2W | average DSC (%): 80.25 -> 81.33 | Helped |
| Apply PSTUDA rather than MUNIT | Kidney tumor segmentation using T1c→T2W adaptation · MSKT: T1c→T2W | average DSC (%): 62.06 -> 65.52 | Helped |
| Apply PSTUDA rather than CycleGAN | Kidney tumor segmentation using T1c→DWI adaptation · MSKT: T1c→DWI | average DSC (%): 79.74 -> 80.61 | Helped |
| Apply PSTUDA rather than CycleGAN | Kidney tumor segmentation using CT→DWI adaptation · KiTS19→MSKT: CT→DWI | average DSC (%): 84.40 -> 82.54 | Hurt |
| Add Progressive Style Transfer (PST) (ablation) | StarGAN v2 baseline · MSKT: T1c→FS T2W | average DSC (%): 66.71 -> 81.34 | Helped |
| Add Multi-Scale Discriminator (MS_D) (ablation) | StarGAN v2 baseline with PST · MSKT: T1c→T2W | average DSC (%): 61.48 -> 65.52 | Helped |
| Replace Mapping Network and Style Encoder with M_SD (ablation) | StarGAN v2 baseline · MSKT: T1c→FS T2W | average DSC (%): 66.71 -> 77.62 | Helped |
| Add Point-wise Instance Normalization (PIN) (ablation) | Baseline with M_SD · MSKT: T1c→T2W | average DSC (%): 48.14 -> 61.48 | Helped |
| Replace PIN with AdaIN (ablation) | PSTUDA · MSKT: T1c→FS T2W | average DSC (%): 81.33 -> 73.58 | Hurt |
| Replace PIN with BIN (ablation) | PSTUDA · MSKT: T1c→T2W | average DSC (%): 65.52 -> 61.58 | Hurt |
| Set M_SD depth to 16384 instead of 4096 (ablation) | PSTUDA · MSKT: T1c→FS T2W | average DSC (%): 81.33 -> 75.69 | Hurt |
| Set M_SD depth to 16384 instead of 4096 (ablation) | PSTUDA · MSKT: T1c→T2W | average DSC (%): 65.52 -> 67.56 | Helped |
Surprising findings
- CycleGAN achieves higher average DSC than PSTUDA on DWI in the CT→MRI setting (by about 1.9%)
- DEPL collapses on T2W in the MSKT experiment (kidney DSC 23.44%), showing structural distortion
Limitations
- Evaluated only on kidney and tumor segmentation with MR sequences; generalization to other tasks is untested
- Framework still operates as a multi-domain to multi-domain model rather than a specialized one-to-multiple model
- Private MSKT dataset is only partially released (checklist says a portion of the data is open-sourced)
Experimental setup
Datasets: MSKT KiTS19 abdominal multi-organ MR/CT dataset
Benchmarks: MSKT KiTS19
Metrics: DSC HD95
Baselines: CycleGAN MUNIT SIFA DEPL StarGAN v2 Supervised training W/o adaptation
Builds on: StarGAN v2 CycleGAN AdaIN U-Net
Closest papers by meaning
- MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.87)
- PrPSeg: Universal Proposition Learning for Panoramic Renal Pathology Segmentation (CVPR 2024 · similarity 0.87)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.86)
- NaMa: Neighbor-Aware Multi-Modal Adaptive Learning for Prostate Tumor Segmentation on Anisotropic MR Images (AAAI 2024 · similarity 0.86)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.86)
- Leveraging Anatomical Consistency for Multi-Object Detection in Ultrasound Images via Source-free Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.86)
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.