CFNet

CFNet encodes rotation equivariance and rotational invariance in convolutional neural networks for analysis of high-throughput microscopy images, enabling automated cell phenotyping and subcellular protein localization classification.

Key Features:

• Rotation Equivariant Conic Convolution: Uses a conic convolution scheme to encode local rotation equivariance within neural network feature extraction.
• Global Rotational Invariance via 2D-DFT: Applies the 2D-discrete-Fourier transform (2D-DFT) to encode global rotational invariance across image orientations.
• Integration within CNN Architectures: Combines conic convolution and 2D-DFT within convolutional neural networks to jointly capture local equivariance and global invariance.
• Comparative Evaluation: Evaluated against standard CNNs and group-equivariant CNNs (G-CNNs) on simulated and real microscopy images focusing on subcellular protein localization.

Scientific Applications:

• Automated Cell Phenotyping: Supports orientation-robust classification of cellular phenotypes from microscopy image datasets.
• Subcellular Protein Localization Classification: Improves classification of protein localization patterns regardless of cell or organelle orientation.
• High-Throughput Microscopy Image Analysis: Applicable to large-scale microscopy datasets requiring rotationally consistent feature recognition.
• Cellular and Tissue Studies: Enables analyses relevant to cellular biology, pathology, and developmental biology that depend on orientation-invariant image interpretation.

Methodology:

CFNet implements conic convolution for local rotation equivariance and applies the 2D-discrete-Fourier transform (2D-DFT) for global rotational invariance within convolutional neural network architectures, and it was compared quantitatively to standard CNNs and group-equivariant CNNs on simulated and real microscopy images targeting subcellular protein localization.