Related Experiment Video
Updated: Aug 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Improving Data-Efficiency and Robustness of Medical Imaging Segmentation Using Inpainting-Based Self-Supervised
Jeffrey Dominic1, Nandita Bhaskhar2, Arjun D Desai1,2
1Department of Radiology, Stanford University, Stanford, CA 94305, USA.
Bioengineering (Basel, Switzerland)
|February 25, 2023
Summary
Self-supervised learning (SSL) with context restoration pretraining improves medical image segmentation accuracy, especially in label-limited settings. This approach enhances model robustness and reduces errors compared to traditional supervised learning.
Area of Science:
- Medical imaging analysis
- Machine learning in healthcare
- Computer vision for diagnostics
Background:
- Self-supervised learning (SSL) offers a promising avenue for medical image segmentation, particularly when labeled data is scarce.
- Inpainting-based pretext tasks are effective for training models without manual annotations.
- Evaluating different SSL training methodologies is crucial for optimizing performance in clinical applications.
Purpose of the Study:
- To systematically evaluate the efficacy of inpainting-based pretext tasks (context prediction and context restoration) for medical image segmentation using SSL.
- To determine the impact of various design choices on the performance of self-supervised U-Net models.
- To compare the performance of optimized SSL models against baseline supervised models in label-limited scenarios.
Main Methods:
- Trained multiple self-supervised U-Net models on MRI and CT datasets using different combinations of design choices and pretext tasks.
- Identified optimal design choices, including context restoration with 32x32 patches and Poisson-disc sampling, transferring only encoder weights, and fine-tuning with a 1e-3 learning rate.
- Compared SSL models with baseline supervised models on clinically-relevant metrics in label-limited conditions.
Main Results:
- SSL pretraining with context restoration, specific patch settings, encoder weight transfer, and fine-tuning parameters significantly improved MRI and CT tissue segmentation accuracy (p < 0.001) over supervised learning.
- Increased unlabeled pretraining data size consistently enhanced segmentation performance across datasets and label-limited scenarios.
- SSL models outperformed supervised models in clinically-relevant metrics, particularly when supervised learning performance was low.
Conclusions:
- SSL pretraining using inpainting-based pretext tasks enhances the robustness of medical image segmentation models in label-limited scenarios.
- This approach effectively reduces worst-case errors often encountered with purely supervised learning methods.
- The findings highlight the potential of SSL to improve diagnostic accuracy and reliability in resource-constrained medical imaging environments.
