An Interpretable 3D Bag-Of-Visual-Words Pipeline for Volumetric Microscopy Classification.

Area of Science:

• Computational biology
• Bioimage analysis
• Microscopy data processing

Background:

• Volumetric microscopy generates complex datasets challenging for traditional analysis.
• Hand-crafted measurements struggle with 3D distributed signals.
• Need for automated, interpretable methods for volumetric data.

Purpose of the Study:

• To develop an interpretable 3D Bag-of-Visual-Words (BoVW) pipeline for classification and analysis of volumetric microscopy data.
• To enable extraction of biologically meaningful structures from 3D datasets.
• To provide localized insights through attention mapping.

Main Methods:

• Utilized a 3D BoVW pipeline detecting multiscale keypoints and rotationally robust descriptors.
• Aggregated features into image-level representations for visualization and classification.
• Employed logistic regression and mapped model weights for attention map generation.
• Applied to cerebellar granule neuron datasets (lattice light-sheet and confocal timelapse).

Main Results:

• Successfully classified nuclei based on chromatin organization differences (control vs. NIPBL loss-of-function).
• Identified fragmented high-attention regions and smoother textures in loss-of-function nuclei.
• Captured ligand-driven receptor clustering and subtle differences in neuronal cultures.
• Demonstrated robustness across ideal and non-ideal imaging conditions.

Conclusions:

• The 3D BoVW pipeline is a simple, interpretable, and broadly applicable framework.
• It effectively extracts biologically meaningful structure from volumetric microscopy data.
• Preserves native 3D context while facilitating classification and analysis.