Unsupervised convolutional autoencoders for 4D transperineal ultrasound classification.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Data Science

Background:

• 4D Transperineal ultrasound (TPUS) is a key imaging modality for assessing female pelvic floor disorders.
• Understanding muscle dynamics during contractions and Valsalva maneuvers is crucial for diagnosis and treatment.
• Current analysis methods may not fully leverage the rich data captured by TPUS.

Purpose of the Study:

• To investigate the feasibility of unsupervised analysis and classification of 4D Transperineal ultrasound (TPUS) data.
• To identify and quantify muscle movements in the female pelvic floor using advanced machine learning techniques.
• To develop a method for objective assessment of pelvic floor function from TPUS imaging.

Main Methods:

• An unsupervised 3D-convolutional autoencoder was employed to compress TPUS frames into latent feature vectors (LFVs).
• Statistical analysis of feature (co)variance was performed to identify movement-specific information.
• Dimensionality reduction (PCA, 2D-autoencoder) and clustering algorithms (k-means, GMM) were applied for data compression and classification.

Main Results:

• Significant differences in features were observed between muscle contraction and Valsalva maneuvers.
• Key features capturing pelvic floor muscle movement were identified through LFVs and their (co)variance.
• Principal component analysis combined with Gaussian mixture models achieved 91.2% accuracy in unsupervised classification of TPUS data.

Conclusions:

• Unsupervised machine learning techniques can effectively extract meaningful information from 4D TPUS data.
• The developed approach provides a robust method for analyzing and classifying pelvic floor muscle activity.
• This study demonstrates the potential of TPUS data for objective and accurate assessment of female pelvic floor disorders.