Related Experiment Video
Updated: Jul 17, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Progressive growing of patch size: Curriculum learning for accelerated and improved medical image segmentation
Stefan M Fischer1, Johannes Kiechle1, Laura Daza2
1School of Computation, Information and Technology, Technical University Munich, Munich, Germany; Department of Radiation Oncology, TUM School of Medicine, TUM University Hospital rechts der Isar, Technical University of Munich, Munich, Germany; Institute of Machine Learning in Biomedical Imaging, Helmholtz Munich, Munich, Germany; Munich Center of Machine Learning (MCML), Munich, Germany.
Progressive Growing of Patch Size (PGPS) enhances 3D medical image segmentation by gradually increasing patch sizes during training. This curriculum learning approach improves segmentation performance and training efficiency across various models and tasks.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Curriculum learning enhances deep learning model training by introducing data complexity progressively.
- 3D medical image segmentation faces challenges with class imbalance and computational costs.
- Existing methods often use fixed patch sizes, limiting training efficiency and performance.
Purpose of the Study:
- Introduce Progressive Growing of Patch Size (PGPS), an automatic curriculum learning strategy for 3D medical image segmentation.
- Evaluate PGPS in resource-efficient and performance-optimized modes to improve segmentation accuracy and reduce training time.
- Demonstrate the broad applicability of PGPS across different architectures and segmentation tasks.
Main Methods:
- Implemented PGPS by starting training with small patch sizes and gradually increasing them.
- Evaluated two modes: resource-efficient (constant batch size) and performance (inversely scaled batch size).
- Assessed segmentation performance (Dice score) and computational costs (FLOPs, wall-clock time) on 15 diverse 3D medical image segmentation tasks.
Main Results:
- The resource-efficient mode matched baseline performance while reducing training time by 56%.
- The performance mode achieved a 1.28% relative gain in mean Dice score, surpassing the baseline across all 15 tasks with 89% of baseline training time.
- PGPS demonstrated significant benefits for tasks with severe foreground-to-background imbalance, like lesion segmentation.
Conclusions:
- PGPS is a broadly applicable and effective strategy for improving 3D medical image segmentation performance and training efficiency.
- The approach consistently boosts performance across diverse models (UNet, UNETR, SwinUNETR) and tasks.
- PGPS reduces segmentation performance variance, making model comparisons more reliable and less sensitive to training stochasticity.