A lightweight neural network for lung nodule detection based on improved ghost module.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Computer tomography (CT) is crucial for lung disease evaluation, but nodule detection is challenging due to image quality and diagnostic workload.
• Deep convolutional neural networks show promise for automated lung nodule detection.
• Current methods require significant physician experience and human-computer interaction.

Purpose of the Study:

• To develop a lightweight neural network for accurate and efficient lung nodule detection.
• To reduce the diagnostic labor for physicians in lung nodule identification.
• To improve the precision and recall of intelligent lung nodule detection systems.

Main Methods:

• A modified GhostNet architecture (Yolov4-GNet) was proposed, incorporating improved bneck structures from MobileNetV3.
• Channel attention and spatial-temporal attention mechanisms were introduced and adjusted.
• Depth-separable convolutions replaced standard 3x3 convolutions to reduce model parameters and enhance network applicability.

Main Results:

• The Yolov4-GNet achieved an F1-score of 0.87, with 86.34% precision and 86.69% recall.
• The proposed network outperformed existing neural networks in precision, recall, and F1-score on the study's dataset.
• Parameter count was significantly reduced compared to original network structures.

Conclusions:

• The developed lung nodule detection method simplifies image processing and improves detection rates and accuracy.
• This research offers a novel and effective approach for lung nodule detection using deep learning.
• The lightweight neural network provides a promising tool for enhancing preoperative lung disease evaluation.