Classification of osteoarthritic and healthy cartilage using deep learning with Raman spectra

Area of Science:

• Biomedical Engineering
• Spectroscopy
• Artificial Intelligence

Background:

• Raman spectroscopy offers rapid molecular analysis of biological tissues.
• Spectral data noise requires extensive pre-processing for accurate analysis.
• Automating pre-processing can enhance the clinical utility of Raman spectroscopy.

Purpose of the Study:

• To develop an end-to-end Convolutional Neural Network (CNN) for automated Raman spectral pre-processing and analysis.
• To classify Raman spectra from superficial and deep cartilage layers in osteoarthritis and osteoporosis patients.
• To identify biologically relevant spectral features using Integrated Gradients.

Main Methods:

• An end-to-end Multi-Convolutional Neural Network (M-CNN) was designed to learn optimal pre-processing strategies.
• The M-CNN was trained and validated using 6-fold cross-validation on cartilage Raman spectra from 45 Osteoarthritis and 19 Osteoporosis patients.
• Integrated Gradients were used to identify key Raman spectral features contributing to network decisions.

Main Results:

• The M-CNN achieved classification accuracy comparable or superior to traditional CNNs on raw or manually pre-processed spectra.
• Identified Raman signatures (features) were confirmed to be biologically relevant.
• Feature selection using Artificial Neural Networks, Decision Trees, and Support Vector Machines demonstrated that minimal features (<3 for disease classification, <300 for layer assignment) yielded comparable performance.

Conclusions:

• The proposed AI approach automates complex pre-processing and feature selection for Raman spectroscopy.
• This method shows potential for facilitating clinical translation of Raman spectroscopy-based diagnostics.
• The technique reduces the need for laborious manual pre-processing and feature selection in spectral analysis.