LungSeek: 3D Selective Kernel residual network for pulmonary nodule diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early detection of pulmonary nodules is crucial for improving lung cancer patient survival rates.
• Current diagnostic methods can be time-consuming and may lack precision.

Purpose of the Study:

• To develop an automated system, LungSeek, for accurate pulmonary nodule detection and classification.
• To enhance the early diagnosis of lung cancer using deep learning.

Main Methods:

• LungSeek employs a two-module system: nodule detection and nodule classification.
• Utilizes a 3D Selective Kernel residual network (SK-ResNet) for nodule detection and a multi-scale feature fusion network for classification.
• The SK-Net module adaptively adjusts receptive fields to learn nodule features effectively.

Main Results:

• The system achieved high accuracy on the luna16 dataset, with false positive rates of 89.06%, 94.53%, and 97.72% at 1, 2, and 4 false positives, respectively.
• LungSeek outperformed state-of-the-art methods, similar networks, and experienced doctors in performance.
• Demonstrated effectiveness in detecting nodules of various sizes.

Conclusions:

• LungSeek provides an effective framework for pulmonary nodule detection and classification from CT scans.
• The proposed 3D SK-ResNet based system shows significant potential for improving early lung cancer diagnosis.