Related Experiment Video
Updated: Jun 10, 2025
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
Less is More: Selective reduction of CT data for self-supervised pre-training of deep learning models with
Daniel Wolf1, Tristan Payer2, Catharina Silvia Lisson3
1Visual Computing Research Group, Institute of Media Informatics, Ulm University, James-Franck-Ring, Ulm, 89081, Germany; Experimental Radiology Research Group, Department for Diagnostic and Interventional Radiology, Ulm University Medical Center, Albert Einstein Allee, Ulm, 89081, Germany.
Reducing redundancy in medical imaging datasets significantly boosts contrastive learning performance and speeds up pre-training. This approach enhances deep learning models for medical image analysis and classification tasks.
Area of Science:
- Medical Imaging
- Deep Learning
- Computer Vision
Background:
- Self-supervised learning with contrastive methods is prevalent in image analysis.
- Contrastive pre-training shows promise for medical images but requires adaptation.
- Medical image characteristics pose unique challenges for current contrastive learning techniques.
Purpose of the Study:
- Investigate how medical image similarity impacts contrastive learning.
- Develop and evaluate strategies to reduce dataset redundancy for improved pre-training.
- Enhance the effectiveness of contrastive pre-training on medical datasets.
Main Methods:
- Explored strategies based on deep embedding, information theory, and hashing.
- Identified and reduced redundancy within medical pre-training datasets.
- Evaluated reduction strategies on contrastive learning across multiple downstream classification tasks.
Main Results:
- Dataset reduction consistently improved downstream task performance (e.g., AUC increased from 0.78 to 0.83 for COVID CT).
- Significant performance gains observed across various classification challenges.
- Pre-training efficiency increased, with speedups up to nine times faster.
Conclusions:
- Dataset quality is crucial for successful contrastive pre-training in medical imaging.
- The proposed method offers a transferable approach to enhance medical image classification.
- Optimizing pre-training datasets improves deep learning model performance on downstream tasks.
