Vibroarthrography-based Knee Lesions Location via Multi-Label Embedding Learning.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• Vibration arthrography (VAG) is a non-invasive, radiation-free technique for knee pathology recognition.
• Current research primarily focuses on knee health status, with limited exploration of VAG for precise lesion localization.

Purpose of the Study:

• To develop an effective method for locating multiple knee lesions simultaneously using VAG signals.
• To overcome limitations of existing Multi-Label classification (MLC) methods in handling label imbalance and sparse correlations for knee lesion identification.

Main Methods:

• Proposed a novel pre-trained model incorporating a label autoencoder (PTM-LAE).
• Utilized a pre-trained feature mapping model with focal loss to address positive-negative label imbalance.
• Employed a Factorization-Machine-based neural network (DeepFM) to extract sparse label correlations between different lesions.

Main Results:

• The PTM-LAE model demonstrated superior performance compared to state-of-the-art methods in knee lesion localization.
• Effectively mitigated issues of positive-negative label imbalance and improved extraction of sparse label correlations.
• Experimental validation on collected VAG data confirmed the model's efficacy.

Conclusions:

• The proposed PTM-LAE method offers a significant advancement in utilizing VAG signals for accurate knee lesion localization.
• This approach has the potential to greatly aid physicians in diagnosis and patient monitoring.
• The study highlights the importance of addressing label imbalance and correlation extraction in medical image analysis.