YOLOv11-MFF: A multi-scale frequency-adaptive fusion network for enhanced CXR anomaly detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Chest X-rays (CXRs) are crucial for diagnosing thoracic diseases, but detecting subtle lesions remains challenging.
• Existing computer-aided diagnosis (CADx) methods struggle with lesions exhibiting ambiguous boundaries, low pixel occupancy, and weak contrast, often overlooking background noise and morphological variations.

Purpose of the Study:

• To develop an enhanced YOLOv11 network, termed YOLOv11-MFF, to address the limitations in detecting subtle and complex lesions in chest X-rays.
• To improve the accuracy and robustness of AI-based anomaly detection in CXRs.

Main Methods:

• Introduced a novel Frequency-Adaptive Hybrid Gate (FAHG) to enhance contrast differentiation between lesions and background.
• Integrated a Multi Scale Parallel Large Convolution (MSPLC) block to expand receptive fields and improve long-range dependency modeling.
• Incorporated a Feature Fusion (FF) module for channel-wise modulation and recalibration to reinforce target-relevant features.

Main Results:

• The YOLOv11-MFF network demonstrated significant improvements in detecting multi-scale and overlapping lesions.
• Achieved superior performance compared to state-of-the-art models on the VinDr-CXR dataset.
• Reported a precision of 48.2%, recall of 42.5%, mAP@0.5 of 41.5%, and mAP@0.5:0.95 of 22.6%.

Conclusions:

• YOLOv11-MFF effectively addresses challenges in chest X-ray anomaly detection, particularly for subtle and complex lesions.
• The proposed innovations in feature extraction and fusion contribute to enhanced diagnostic accuracy in thoracic imaging.
• This advanced AI model shows significant potential for improving computer-aided diagnosis in clinical practice.