Quantification of anomalies in rats' spinal cords using autoencoders

Area of Science:

• Neuroimaging
• Medical Diagnostics
• Computational Biology

Background:

• Computed tomography (CT) and magnetic resonance imaging (MRI) are standard for spinal cord lesion evaluation.
• Accurate quantification of spinal cord lesions is crucial for understanding disease progression.
• Semi-supervised anomaly detection is well-suited for spinal cord analysis due to limited labeled data.

Purpose of the Study:

• To develop and evaluate autoencoder and variational autoencoder models for quantifying spinal cord lesions.
• To generate lesion progression data through aggregated anomaly-based scores.
• To explore complementary quantification methods for enhanced diagnostic accuracy.

Main Methods:

• Analysis of micro-CT scans of rat spinal cords.
• Application of semi-supervised, deviation-based anomaly detection algorithms.
• Large-scale evaluation of autoencoders and variational autoencoders for lesion quantification.
• Introduction of an area-based model for lesion quantification.

Main Results:

• Several autoencoder models successfully generated 3D spinal cord lesion quantifications.
• One model achieved an average correlation of 0.83 with weakly labeled true data.
• An area-based model demonstrated a mean correlation of 0.84.
• The complementary use of autoencoder-based methods and area features was discussed.

Conclusions:

• Autoencoder-based methods provide a powerful approach for spinal cord lesion quantification.
• The developed methods show high correlation with existing data, promising for medical diagnostics.
• Future applications include lesion clustering and improved understanding of disease progression.