Self-supervised learning improves robustness of deep learning lung tumor segmentation models to CT imaging...

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computer Vision

Background:

• Self-supervised learning (SSL) extracts features from unlabeled data for downstream tasks.
• Self-pretraining uses curated datasets for both pretraining and fine-tuning.
• Uncurated public medical images offer potential for robust foundation models via 'wild' SSL.

Purpose of the Study:

• Compare robustness of wild vs. self-pretrained models (CNN, ViT, Swin) for non-small cell lung cancer (NSCLC) segmentation.
• Evaluate models on 3D computed tomography (CT) scans.
• Assess impact of different pretraining strategies on model performance.

Main Methods:

• Wild-pretraining and self-pretraining applied to CNN, ViT, and Swin models using unlabeled and curated NSCLC CT datasets.
• Utilized masked image transformer pretext tasks for both pretraining approaches.
• Fine-tuned and tested models on diverse NSCLC datasets, evaluating accuracy, robustness to imaging variations, and feature reuse.

Main Results:

• Wild-pretrained Swin models exhibited higher feature reuse and outperformed self-pretrained models.
• ViT and CNN models did not show a significant benefit from wild-pretraining over self-pretraining.
• Masked image prediction pretext task yielded higher accuracy than contrastive tasks.

Conclusions:

• Wild-pretrained Swin networks demonstrated superior robustness for lung tumor segmentation against CT imaging variations.
• ViT and CNN models did not gain a clear advantage from wild-pretraining compared to self-pretraining.
• The choice of architecture and pretraining strategy impacts robustness in medical image analysis.