SPCF-YOLO: An Efficient Feature Optimization Model for Real-Time Lung Nodule Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate pulmonary nodule detection in CT scans is crucial for early lung cancer diagnosis.
• Conventional deep learning models struggle with fragmented feature integration, limiting detection accuracy.

Purpose of the Study:

• To introduce SPCF-YOLO, a real-time deep learning framework for enhanced pulmonary nodule detection.
• To improve the integration of hierarchical features and anatomical context in CT imaging analysis.

Main Methods:

• The proposed SPCF-YOLO framework utilizes space-to-depth convolution (SPDConv) to preserve fine-grained features.
• It incorporates shared feature pyramid convolution (SFPConv) for dynamic multi-scale contextual information extraction.
• Enhanced attention modules (PSA and CoTB) and a small object detection layer improve sensitivity and reduce feature loss.

Main Results:

• SPCF-YOLO achieved 82.8% mAP and 82.9% F1 score on the LUNA16 dataset.
• The framework operates at 151 frames per second, demonstrating real-time performance.
• Significant improvements of 17.5% and 82.9% over YOLOv8 were observed.

Conclusions:

• SPCF-YOLO offers a robust and efficient solution for real-time pulmonary nodule detection in CT imaging.
• The framework shows potential for clinical viability and improved diagnostic accuracy.
• Cross-modality validation confirmed the model's strong generalization capabilities.