Related Experiment Video
Updated: Aug 30, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
STAMP: Simultaneous Training and Model Pruning for low data regimes in medical image segmentation.
Nicola K Dinsdale1, Mark Jenkinson2, Ana I L Namburete3
1Wellcome Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, UK; Oxford Machine Learning in NeuroImaging Lab (OMNI), Department of Computer Science, University of Oxford, UK.
We developed the STAMP algorithm for efficient medical image segmentation, enabling robust UNet training with targeted dropout. This method significantly reduces network size while enhancing performance, even with limited labeled data.
Area of Science:
- Medical image analysis
- Machine learning
- Computer vision
Background:
- High-quality manual annotations are crucial for developing medical image segmentation algorithms.
- Training deep learning models like convolutional neural networks requires extensive expert time, knowledge, and large labeled datasets.
- Current methods face challenges in achieving high performance with limited available annotations.
Purpose of the Study:
- To develop an efficient algorithm for simultaneous training and pruning of UNet architectures for medical image segmentation.
- To enhance network robustness to pruning through targeted channelwise dropout.
- To enable effective training of segmentation models even with scarce labeled data.
Main Methods:
- Introduced the STAMP (Simultaneous Training and Pruning) algorithm.
- Implemented targeted channelwise dropout within a UNet architecture.
- Validated the technique across diverse segmentation tasks and medical imaging modalities.
Main Results:
- Achieved significant reduction in network size (over 85% parameter reduction) compared to standard UNet models.
- Demonstrated substantially higher performance than equivalent standard UNet models.
- Showcased the algorithm's effectiveness across various segmentation tasks and imaging types.
Conclusions:
- The STAMP algorithm allows for the creation of highly performant and significantly smaller medical image segmentation networks.
- Online pruning with targeted dropout enhances model robustness and efficiency.
- This approach holds potential for training networks on datasets with very limited labeled data, addressing a key bottleneck in medical AI development.
