YOLO-ED: An efficient lung cancer detection model based on improved YOLOv8.

Area of Science:

• Medical imaging
• Computer vision
• Artificial intelligence

Background:

• You Only Look Once (YOLO) models offer scalability and generalization for medical object detection.
• Existing YOLO models face challenges with computational cost and precision due to consecutive convolutions and bilinear interpolation.

Purpose of the Study:

• To propose an improved YOLOv8-based model (YOLO-ED) for enhanced lung cancer detection in CT images.
• To address limitations of existing models by reducing computational overhead and improving detection accuracy.

Main Methods:

• Developed YOLO-ED by integrating an Efficient Modulation module for weighted feature fusion and a DySample module for dynamic upsampling.
• The Efficient Modulation module reduces model parameters and computational load.
• The DySample module replaces conventional upsampling to minimize information loss and improve feature extraction accuracy.

Main Results:

• YOLO-ED demonstrated significant improvements in precision across lung cancer detection tasks.
• The model achieved a notable reduction in computational cost compared to baseline models.
• Validation was performed on the LUNG-PET-CT-DX and LUNA16 datasets, confirming YOLO-ED's effectiveness.

Conclusions:

• YOLO-ED offers a superior approach for medical image object detection, specifically for lung cancer in CT scans.
• The proposed model effectively balances precision and computational efficiency.
• YOLO-ED represents a significant advancement in applying deep learning for medical image analysis.