Enhancing Lung Nodule Classification with Triple Union Efficient Attention and Neural Architecture Search

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lung nodule classification is crucial for early lung cancer screening and treatment decisions.
• Existing methods face challenges with large model sizes and lack of spatial/channel-aware information, limiting clinical application and interpretability.

Purpose of the Study:

• To develop an efficient and interpretable deep learning model for lung nodule classification.
• To address limitations of existing models in terms of size, resource requirements, and feature representation.

Main Methods:

• Proposed TripleAttentionNet with Triple Union Efficient Attention (TUEA) modules for feature extraction.
• Implemented weight sharing across attention pathways to reduce parameters and enhance correlations.
• Utilized Neural Architecture Search (NAS) to optimize network architecture for accuracy and size.

Main Results:

• TripleAttentionNet achieved state-of-the-art performance on LIDC-IDRI, LUNA16, and LUNGx datasets.
• The model demonstrated effective capture of key lung nodule features with linear complexity.
• Achieved a favorable balance between classification accuracy and model parameter size.

Conclusions:

• TripleAttentionNet offers an effective solution for lung nodule classification in early cancer screening.
• The model's efficiency and interpretability make it suitable for resource-limited clinical settings.
• This approach has the potential to enhance computer-aided diagnosis in lung cancer screening.