Automatic Joint Lesion Detection by enhancing local feature interaction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning excels in Automatic Joint Lesion Detection (AJLD), but balancing accuracy and efficiency is challenging.
• Clinical requirements necessitate high accuracy in end-to-end lesion detection for conditions like arthritis.

Purpose of the Study:

• To enhance the accuracy and efficiency of Automatic Joint Lesion Detection (AJLD) for clinical applications.
• To introduce novel modules for improving deep learning models in detecting joint lesions, particularly in arthritis.

Main Methods:

• Integration of Local Attention Feature Fusion (LAFF) and Gaussian Positional Encoding (GPE) modules into YOLO models.
• Development of an improved YOLO model, YOLO_lf, focusing on enhanced local feature interaction.
• Validation on multiple datasets, including large-scale arthritis X-ray datasets and the MURA dataset.

Main Results:

• The proposed YOLO_lf model demonstrated significant increases in detection accuracy across various YOLO architectures.
• YOLO_lfv8 showed improved mAP@50 from 0.765 to 0.785 and 0.831 to 0.859 on two arthritis datasets.
• The LAFF and GPE modules proved to be plug-and-play, enhancing performance without architectural overhauls.

Conclusions:

• The novel LAFF and GPE modules effectively enhance deep learning models for AJLD, particularly for arthritis detection.
• The improved YOLO_lf models offer a promising, clinically applicable solution for accurate and efficient joint lesion detection.