Unsupervised segmentation of biomedical hyperspectral image data: tackling high dimensionality with convolutional...

Area of Science:

• Biomedical optics
• Medical imaging analysis
• Computational pathology

Background:

• Biopsy specimen analysis is crucial for disease monitoring and diagnosis.
• Hyperspectral imaging (HSI) captures spatial and spectral information of molecular arrangements in tissues.
• High dimensionality of HSI data presents challenges for extracting and utilizing features for image segmentation.

Purpose of the Study:

• To develop and evaluate an end-to-end deep learning approach for spatio-spectral segmentation of HSI data.
• To compare the performance of end-to-end trained convolutional autoencoders (CAEs) against non-end-to-end and spectral-only methods.
• To assess the utility of different CAE architectures for segmenting biomedical HSI data.

Main Methods:

• Utilized deep convolutional autoencoders (CAEs) for end-to-end feature extraction and spatio-spectral clustering of HSI patches.
• Compared end-to-end CAE segmentation with non-end-to-end training and spectral k-means on simulated porcine tissue HSIs.
• Evaluated three CAE architectures (2D CAE, 3D CAE, UwU-net inspired) on IR HSIs of real colon samples.

Main Results:

• The end-to-end spatio-spectral segmentation approach using CAEs demonstrated robust performance.
• All tested CAE architectures produced segmentations that corresponded well with HE stained adjacent tissue slices.
• The study highlights the potential of CAE-driven spatio-spectral clustering for biomedical HSI segmentation.

Conclusions:

• Deep convolutional autoencoders offer a powerful tool for end-to-end spatio-spectral segmentation of biomedical hyperspectral images.
• The proposed method shows promise for improving disease diagnosis and monitoring through detailed tissue analysis.
• Further research with more accurate ground truth data is needed to precisely compare different CAE architectures.