An attention aided wavelet convolutional neural network for lung nodule characterization

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a major cause of cancer mortality globally.
• Early detection of pulmonary nodules is crucial for improving patient prognosis and survival rates.
• Distinguishing benign from malignant nodules via conventional imaging remains a significant clinical challenge.

Purpose of the Study:

• To propose a novel two-pathway wavelet-based deep learning computer-aided diagnosis (CADx) framework for enhanced lung nodule classification.
• To improve the accuracy and efficiency of lung nodule characterization using high-resolution computed tomography (HRCT) images.

Main Methods:

• Developed the Wavelet-based Lung Cancer Detection Network (WaveLCDNet) utilizing convolutional neural network (CNN) blocks and trainable wavelet blocks for multi-resolution analysis.
• Incorporated a convolutional block attention module (CBAM) to enhance discriminative feature learning.
• Employed adaptive fusion of extracted features followed by global average pooling (GAP).

Main Results:

• WaveLCDNet achieved high performance on the LIDC-IDRI dataset with sensitivity, specificity, and accuracy of 96.89%, 95.52%, and 96.70%, respectively.
• External validation on the Kaggle DSB2017 dataset demonstrated 95.90% accuracy and a Brier Score of 0.0215.
• The framework showed reliability across independent imaging sources, indicating practical value for clinical integration.

Conclusions:

• The proposed framework effectively combines multi-scale convolutional filtering with wavelet-based multi-resolution analysis and attention mechanisms.
• WaveLCDNet outperforms state-of-the-art deep learning models for lung nodule characterization.
• This CADx solution offers a promising approach for enhancing early lung cancer diagnosis in clinical settings.