Multi-scale Lesion Feature Fusion and Location-Aware for Chest Multi-disease Detection.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Radiology

Background:

• Current chest X-ray analysis often focuses on single diseases, overlooking co-occurring conditions.
• Lesion diversity in location and attributes complicates accurate multi-disease detection.
• Existing methods struggle with diminished accuracy when multiple diseases are present.

Purpose of the Study:

• To develop an advanced detection framework for improved multi-disease identification in chest X-rays.
• To enhance lesion feature extraction, fusion, and positional perception for greater diagnostic accuracy.
• To address limitations in current methods that disregard multiple pathologies in a single scan.

Main Methods:

• A multi-scale lesion feature extraction network was designed to capture unique lesion characteristics and locations.
• An instance-aware semantic enhancement network was introduced for adaptive feature fusion across scales.
• Lesion region feature mapping using candidate boxes was employed to preserve critical positional information.

Main Results:

• The proposed framework demonstrated a 6% increase in mean average precision (mAP).
• A significant 8.4% improvement in mean recall (mR) was observed compared to baseline methods.
• Experimental results on the VinDr-CXR dataset validated the model's effectiveness.

Conclusions:

• The novel framework effectively detects multiple chest diseases by accurately capturing specific features and location information.
• Enhanced multi-scale feature extraction and fusion improve lesion perception and diagnostic performance.
• This approach offers a promising solution for more comprehensive chest X-ray analysis.