Similarity maps and hierarchical clustering for annotating FT-IR spectral images

Area of Science:

• Microscopy and Imaging
• Computational Pathology
• Biomedical Data Analysis

Background:

• Unsupervised segmentation is crucial for annotating infrared microscopic images in label-free spectral histopathology.
• Fourier Transform Infrared (FT-IR) microscopic image segmentation relies on diverse clustering approaches for histopathological agreement.

Purpose of the Study:

• To introduce interactive similarity maps as an alternative annotation strategy for infrared microscopic images.
• To compare the segmentation accuracy of interactive similarity maps with conventional hierarchical clustering methods.
• To develop a quantitative validation scheme for hierarchical clustering in infrared microscopy.

Main Methods:

• Introduction of interactive similarity maps for image annotation.
• Quantitative comparison of segmentation accuracy between interactive similarity maps and hierarchical clustering.
• Development of a scheme to identify non-horizontal cuts in dendrograms for validation.

Main Results:

• Segmentations from interactive similarity maps achieve accuracy comparable to hierarchical clustering.
• A novel validation scheme for hierarchical clustering in infrared microscopy was established.
• Hierarchical two-means clustering demonstrates performance on par with Ward's clustering.

Conclusions:

• Interactive similarity maps present a viable and attractive alternative to hierarchical clustering for infrared microscopic image annotation.
• The developed validation scheme confirms the comparable performance of hierarchical two-means and Ward's clustering.
• Hierarchical two-means offers a more efficient, less resource-demanding alternative for annotating large spectral images.