Deep neural networks can differentiate thyroid pathologies on infrared hyperspectral images

Area of Science:

• Biomedical optics
• Digital pathology
• Machine learning in healthcare

Background:

• Thyroid pathologies like thyroiditis, hypothyroidism, and cancer significantly impact health.
• Histological analysis is crucial for diagnosing thyroid diseases, but current methods face limitations.
• Hyperspectral imaging offers a novel approach to analyze biological samples based on molecular light interactions.

Purpose of the Study:

• To develop a method for acquiring infrared absorbance spectra from histological thyroid specimens.
• To create a deep learning model for classifying thyroid tissue based on spectral patterns.
• To differentiate between healthy, cancerous, and goiter thyroid tissues.

Main Methods:

• Acquisition of infrared absorbance spectra from each voxel of histological thyroid specimens.
• Development of a multiclass fully-connected neural network model.
• Utilizing k-fold cross-validation for model performance evaluation.

Main Results:

• Achieved an average accuracy of 93.66% in classifying thyroid tissue.
• Demonstrated high sensitivity (93.47%) and specificity (96.93%).
• Confirmed the feasibility of using infrared hyperspectral imaging for thyroid pathology characterization.

Conclusions:

• Infrared hyperspectral imaging effectively characterizes healthy thyroid tissue and pathologies via absorbance measurements.
• The proposed deep learning model shows potential for enhancing diagnostic efficiency.
• This technology can lead to improved patient outcomes in thyroid disease management.