Self-Supervised Learning Method for 3D Detection of Lung Cancer Based on PET Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Positron Emission Tomography (PET) is crucial for lung cancer detection.
• AI enhances PET-based lung cancer diagnosis but requires extensive labeled data.
• Acquiring sufficient labeled PET imaging data for AI training is a significant challenge.

Purpose of the Study:

• To enhance the accuracy of 3D lung cancer detection in PET images.
• To address the challenge of limited labeled data in AI-driven PET imaging analysis.
• To propose a novel self-supervised learning method utilizing pseudo image generation.

Main Methods:

• A spatial tumor simulator generated 3D pseudo-lesions, implanted into normal PET images to create pseudo-lung-cancer images.
• A self-supervised restoration task using paired original and pseudo-lesion PET images was designed for pretraining.
• A Dual-Attention Hybrid Unet (DH-Unet) with self-attention was pretrained and then fine-tuned on real labeled PET data for 3D lung cancer detection.

Main Results:

• The proposed method demonstrated significant performance in 3D lung cancer detection.
• Achieved an mAP@0.10-0.50 of 0.4616.
• Outperformed random initialization by 13.72% and traditional self-supervised models by 11.1%.

Conclusions:

• The self-supervised learning framework effectively improves lung cancer detection in PET imaging.
• The combination of pseudo image generation and self-attention DH-Unet is particularly beneficial with limited labeled data.
• This approach offers a viable solution for enhancing AI-based lung cancer diagnosis using PET scans.