Single Modality vs. Multimodality: What Works Best for Lung Cancer Screening?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Clinical decisions for cancer diagnosis and treatment often integrate multiple data sources.
• Artificial intelligence (AI) models can enhance diagnostic accuracy by mimicking this multimodal approach.
• Lung cancer, with its high mortality due to late diagnosis, presents a significant opportunity for AI-driven multimodal analysis.

Purpose of the Study:

• To investigate the predictive capability of combining computed tomography (CT) imaging and clinical data for lung cancer detection.
• To develop and compare single-modality and multimodality AI models for lung cancer prediction.
• To explore the potential of fusing different data types for a more comprehensive disease analysis.

Main Methods:

• Utilized the National Lung Screening Trial dataset, including CT scans and clinical data.
• Developed a ResNet18 network for classifying 3D CT nodule regions of interest (ROI).
• Developed a random forest algorithm for classifying clinical data.
• Implemented three multimodality strategies (intermediate and late fusion) to combine imaging and clinical data.

Main Results:

• The ResNet18 model achieved an area under the ROC curve (AUC) of 0.7897 for CT data.
• The random forest model achieved an AUC of 0.5241 for clinical data.
• The best multimodality model, using a fully connected layer to fuse deep imaging features and clinical data, achieved an AUC of 0.8021.

Conclusions:

• Multimodality AI models integrating CT scans and clinical data demonstrate superior performance in lung cancer prediction compared to single-modality models.
• Fusing diverse data sources allows for a more comprehensive analysis, potentially leading to improved diagnostic accuracy for complex diseases like lung cancer.
• This approach highlights the potential of AI to enhance clinical decision-making by leveraging multiple data modalities.