Downsampling attention fusion network (DAFNet): a You Only Look Once network for lung nodule detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate lung nodule detection in CT imaging is vital for early lung disease diagnosis.
• Challenges include varying nodule characteristics and tissue interference.
• Deep learning offers potential for improved detection accuracy and efficiency.

Purpose of the Study:

• To develop a novel deep learning framework, DAFNet, for enhanced lung nodule detection.
• To improve both the accuracy and efficiency of lung nodule detection in CT images.
• To provide a flexible and scalable solution for lung nodule detection.

Main Methods:

• Developed the Downsampling Attention Fusion Network (DAFNet) with a dual-branch downsampling module for multiscale feature extraction.
• Incorporated a global attention module (GAM) to enhance feature representation.
• Evaluated DAFNet on a lung CT image dataset (LCTD) and the LUNA16 dataset using mAP and recall metrics.

Main Results:

• DAFNet achieved 93.2% precision and 90.7% recall on the LCTD.
• Outperformed state-of-the-art methods on LUNA16 in terms of mAP.
• Demonstrated near-real-time inference (1.5 ms/image) with a lightweight model (2.5 M parameters).

Conclusions:

• DAFNet presents a novel and effective deep learning approach for lung nodule detection.
• The proposed modules offer plug-and-play integration, enhancing framework flexibility.
• Further validation is needed for clinical translation, focusing on false-positive reduction and malignancy grading.