Multimodal PET/CT-based PD-L1 status prediction in lung cancer via semi-supervised and unsupervised deep learning

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Deep learning (DL) shows promise in lung cancer screening and prognosis.
• Combining 18FDG PET/CT with DL aids in predicting programmed death ligand-1 (PD-L1) expression, improving accuracy and aiding clinical decisions.
• Acquiring large, high-quality annotated medical datasets for DL is challenging due to expert requirements and regulatory hurdles.

Purpose of the Study:

• To propose a semi-supervised and unsupervised deep neural network (USSLNet) for predicting PD-L1 expression in lung cancer.
• To leverage early fusion multi-modal PET/CT images for improved DL performance.
• To address the challenge of limited labeled medical data in DL models.

Main Methods:

• Developed USSLNet, a semi-supervised and unsupervised deep neural network framework.
• Utilized early fusion of multi-modal PET/CT images.
• Employed alternate task training to propagate label information to unlabeled data, mitigating overfitting.

Main Results:

• The USSLNet framework demonstrated improved robustness and reduced outlier impact compared to existing methods.
• Achieved superior performance in PD-L1 status classification.
• Consistently outperformed current approaches when using various unlabeled PET/CT image types.

Conclusions:

• The proposed USSLNet effectively predicts PD-L1 expression using limited and partially annotated multi-modal PET/CT datasets.
• This approach offers a viable non-invasive imaging biomarker for lung cancer patient stratification.
• The method enhances DL model performance in medical imaging despite data limitations.