Lung Nodule Malignancy Prediction From Longitudinal CT Scans With Siamese Convolutional Attention Networks

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computational pathology

Background:

• Accurate lung nodule malignancy prediction is crucial for early cancer detection.
• Existing methods often require significant computational resources and large datasets.
• Integrating temporal information, like nodule growth, can improve diagnostic accuracy.

Purpose of the Study:

• To develop a novel convolutional attention-based network for enhanced malignancy prediction.
• To leverage pre-trained 2-D convolutional feature extractors for efficiency.
• To evaluate the network's performance in both single-time-point and multi-time-point classification scenarios.

Main Methods:

• A Siamese structure was employed for multi-time-point classification.
• The framework utilizes pre-trained 2-D convolutional neural networks (CNNs) as feature extractors.
• Attention mechanisms were incorporated to focus on relevant image regions.

Main Results:

• The proposed 2-D CNN method outperformed a comparable 3-D network in single-time-point classification.
• The new approach used less than half the parameters of the 3-D network.
• Performance gains were observed in multi-time-point classification, highlighting the value of temporal data.

Conclusions:

• Attention-based, Siamese 2-D pre-trained CNNs offer an effective and efficient solution for malignancy prediction.
• The framework demonstrates fast training times and high accuracy using both single and multiple imaging time points.
• This approach holds promise for improving the clinical diagnosis of lung nodules.