Hyperspectral imaging-based cutaneous wound classification using neighbourhood extraction 3D convolutional neural...

Area of Science:

• Medical Imaging
• Computational Biology
• Dermatology

Background:

• Hyperspectral imaging (HSI) offers advanced capabilities for medical research and clinical applications.
• HSI provides crucial information for characterizing tissue, with distinct spectral signatures for wounded versus normal tissue due to oxygenation changes.

Purpose of the Study:

• To classify cutaneous wounds using hyperspectral imaging and a novel neighborhood extraction 3D convolutional neural network (CNN).
• To leverage spatial and spectral information from HSI data for accurate wound classification.

Main Methods:

• Acquired hyperspectral images of wounded and normal tissues.
• Generated 3D cuboids incorporating neighboring pixels to capture spatial and spectral data.
• Trained a 3D CNN model to classify tissues based on these cuboids.

Main Results:

• The proposed 3D CNN method achieved 99.69% accuracy with a 0.9/0.1 training/testing split and a cuboid spatial dimension of 17.
• The method demonstrated superior performance compared to 2D CNNs, even with limited training data.
• High classification accuracy for wounded areas was consistently observed.

Conclusions:

• Hyperspectral imaging combined with a 3D CNN is a highly effective clinical diagnostic tool for classifying wounded and normal tissues.
• The method's success is independent of skin color, as spectral signatures of wounded and normal tissues remain consistent across different ethnic groups.