Unsupervised Semantic Segmentation Models for Region of Interest Identification

Area of Science:

• Computational pathology
• Bioinformatics
• Image analysis

Background:

• Spatial omics technologies like mass spectrometry imaging (MSI) provide molecular distributions but require correlation with tissue morphology.
• Manual annotation of tissue regions of interest (ROIs) for MSI data is time-consuming.
• Automated segmentation models offer a time-saving alternative for annotating ROIs.

Purpose of the Study:

• To compare the performance of eight unsupervised semantic segmentation algorithms for automated ROI annotation in MSI data.
• To evaluate the impact of noise reduction techniques on segmentation accuracy.
• To provide recommendations for selecting optimal segmentation models for spatial omics studies.

Main Methods:

• Eight unsupervised semantic segmentation algorithms (R and Python) were applied to PAS-stained kidney and plant root images.
• Manual annotations were used as ground truth for performance evaluation.
• Dimension reduction techniques were tested for noise reduction.
• Performance metrics (e.g., balanced accuracy, time) were calculated for each model.

Main Results:

• K-means and pytorch-tip demonstrated the best performance in terms of balanced accuracy and processing time, particularly at smaller cluster sizes.
• All algorithms showed decreased performance with increasing cluster numbers.
• The choice of segmentation model significantly impacted downstream statistical analyses.

Conclusions:

• Unsupervised segmentation models can effectively automate tissue ROI annotation for spatial omics data.
• Model selection is crucial and should be based on case-specific performance evaluation, as no single model excels in all scenarios.
• K-means and pytorch-tip are recommended for specific applications requiring a balance of accuracy and speed.