A Self-Supervised Foundation Model Based on Three-Dimensional Chest CT Scans for Lung Cancer Diagnosis and Prognosis...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiomics and Computational Pathology
• Oncology and Diagnostic Radiology

Background:

• Lung cancer diagnosis and prognosis rely heavily on accurate interpretation of chest CT scans.
• Existing deep learning models often require large labeled datasets, limiting their application.
• Foundation models offer a promising approach for generalizable medical image analysis.

Purpose of the Study:

• To develop and validate a self-supervised foundation model for chest CT analysis in lung cancer.
• To evaluate the model's performance in predicting histologic subtype, cancer stage, survival, and recurrence.
• To compare the proposed model against existing machine learning and deep learning algorithms.

Main Methods:

• Development of the Unified CT-Based Lung Cancer Imaging Foundation (UCLIF) model using self-supervised learning on over 33,000 chest CT scans.
• Pretraining with a contrastive masked image modeling task, followed by fine-tuning on specific lung cancer clinical tasks.
• Performance assessment using accuracy, sensitivity, specificity, and AUC, with superiority testing via the DeLong test.

Main Results:

• The UCLIF model achieved superior performance compared to models pretrained on natural images or single tumor regions (P < .001).
• High AUCs were reported for histologic subtype classification (e.g., 0.96 for adenocarcinoma), cancer staging (e.g., 0.99 for stage II), survival prediction (e.g., 0.97 for 1-year survival), and recurrence prediction (AUC, 0.95).

Conclusions:

• The UCLIF foundation model demonstrates high accuracy in predicting critical lung cancer characteristics from chest CT data.
• Self-supervised learning enables robust performance in diverse clinical tasks without extensive labeled data.
• This AI model holds potential for improving lung cancer diagnosis, staging, and patient outcome prediction.