Vision-language model-based semantic-guided imaging biomarker for lung nodule malignancy prediction.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Oncology
• Computer-Aided Diagnosis

Background:

• Machine learning models for lung nodule malignancy assessment often struggle with interpretability and reliance on manual annotation.
• Existing models face challenges in robustness due to sensitivity to imaging variations in clinical settings.

Purpose of the Study:

• To develop a machine learning model that integrates semantic features from radiologists to predict lung cancer.
• To enhance the clinical relevance, robustness, and explainability of lung nodule malignancy assessment.

Main Methods:

• Utilized 938 low-dose CT scans (NLST) and 2,625 lesions (LIDC) with semantic features; external datasets from UCLA, LUNGx, and Duke were included.
• Converted structured semantic features into text and fine-tuned a pretrained Contrastive Language-Image Pre-training (CLIP) model using parameter-efficient fine-tuning.
• Predicted one-year lung cancer diagnosis by aligning imaging and semantic text features.

Main Results:

• Achieved an AUROC of 0.901 and AUPRC of 0.776 on the NLST test set, outperforming SOTA models.
• Demonstrated robust performance across multiple external validation datasets.
• Zero-shot inference using CLIP yielded high AUROCs for nodule margin (0.807), consistency (0.812), and pleural attachment (0.840).

Conclusions:

• The proposed vision-language model integrating semantic features surpasses SOTA models in lung cancer prediction from diverse CT scans.
• The approach offers explainable outputs, assisting clinicians in understanding model predictions.
• The study provides a pathway for more robust and interpretable AI in lung cancer screening.