Research map / Detection and segmentation / Open-set and few-shot detection
Learn2Synth: Learning Optimal Data Synthesis Using Hypergradients for Brain Image Segmentation
Xiaoling Hu, Xiangrui Zeng, Oula Puonti, Juan Eugenio Iglesias, Bruce Fischl, Yaël Balbastre
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn augmentation parameters with hypergradients (Learn2Synth, parametric fixed sigma) | Naive SynthSeg · OASIS3 (2D) | Dice: 0.831 -> 0.860 | Improved |
| learn augmentation parameters with hypergradients (Learn2Synth, parametric fixed sigma) | Naive SynthSeg · ABIDE (2D) | Dice: 0.869 -> 0.878 | Improved |
| use nonparametric UNet augmentation (fixed sigma) instead of naive synthesis | Naive SynthSeg · OASIS3 (2D) | Dice: 0.831 -> 0.875 | Improved |
| Learn2Synth instead of mixing real images into SynthSeg training | Mixed SynthSeg · OASIS3 (2D) | Dice: 0.854 -> 0.875 | Improved |
| fine-tune SynthSeg on real images versus Learn2Synth | Finetuned SynthSeg · ABIDE (2D) | Dice: 0.871 -> 0.874 | No clear effect |
| replace accurate parametric noise+INU model with nonparametric model (ablation) | Learn2Synth noise+INU · synthetic noise + INU, train sigma=0 test sigma=0 | Dice: 0.924 -> 0.903 | Hurt |
| use nnUNet instead of UNet backbone for Learn2Synth (ablation) | Learn2Synth (parametric, fixed sigma) · ABIDE (2D) | Dice: 0.878 -> 0.893 | Improved |
| use model selection by validation set for supervised UNet (ablation) | Supervised UNet (29 train) · FLASH 3deg | Dice: 0.301 -> 0.419 | Improved |
| use validation set for model selection in Learn2Synth (ablation) | Learn2Synth (29 train) · FLASH 3deg | Dice: 0.801 -> 0.804 | No clear effect |
| reduce Learn2Synth training set from 29 to 1 sample (ablation) | Learn2Synth · Buckner39 test (MPRAGE) | Dice: 0.895 -> 0.718 | Hurt |
| Learn2Synth instead of SynthSeg on unseen FLASH contrast | SynthSeg · FLASH 5deg | Dice: 0.694 -> 0.789 | Improved |
| Learn2Synth instead of supervised UNet on unseen FLASH contrast | Supervised UNet · FLASH 3deg | Dice: 0.419 -> 0.804 | Improved |
Closest papers by meaning
- Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation (CVPR 2018 · similarity 0.88)
- Semantic Segmentation With Generative Models: Semi-Supervised Learning and Strong Out-of-Domain Generalization (CVPR 2021 · similarity 0.88)
- Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation (CVPR 2019 · similarity 0.87)
- Automated Synthetic-to-Real Generalization (ICML 2020 · similarity 0.87)
- Factorized Diffusion Architectures for Unsupervised Image Generation and Segmentation (NeurIPS 2024 · similarity 0.86)
- Real-Fake: Effective Training Data Synthesis Through Distribution Matching (ICLR 2024 · 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.
