Lung Nodule Malignancy Classification On 3D CT Images Using a Cosine Similarity-Enhanced Graph Attention Network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer diagnosis at advanced stages contributes to high mortality rates.
• Accurate pulmonary nodule classification is crucial for early detection and intervention.
• Traditional methods for nodule analysis have limitations, including high false-positive rates and reliance on individual CT slices.

Purpose of the Study:

• To develop an end-to-end patient-level classification framework for pulmonary nodule malignancy.
• To address the limitations of existing methods by integrating spatial and contextual information.
• To improve diagnostic accuracy and reduce false positives in pulmonary nodule classification.

Main Methods:

• Utilized a pre-trained Vision Transformer (ViT) for slice-level feature extraction.
• Constructed patient graphs using spatial adjacency and cosine similarity for edge connections.
• Employed a Cosine Similarity-Enhanced Graph Attention Network (CSEGAT) for patient-level classification, incorporating a similarity-enhanced attention mechanism.

Main Results:

• The CSEGAT model achieved a peak testing accuracy of 90.85%.
• Demonstrated average sensitivity of 88.84% and average specificity of 90.65%.
• Outperformed state-of-the-art methods on the LIDC-IDRI dataset in accuracy, sensitivity, specificity, and AUC.

Conclusions:

• The proposed CSEGAT framework offers a robust solution for pulmonary nodule diagnosis.
• End-to-end learning from slice features to patient diagnosis enhances accuracy and reduces false positives.
• The method effectively captures spatial and contextual relationships for improved classification performance.