mi-CNN

mi-CNN classifies subtissue regions in mass spectrometry imaging (MSI) data using a multiple-instance convolutional neural network to enable subtissue-level molecular discrimination from weakly labeled tissue-level annotations.

Key Features:

• Multiple Instance Learning (MIL): Leverages MIL to enable weak supervision from tissue-level annotations, allowing the use of MSI data when subtissue-level ground truth is unavailable.
• Convolutional Neural Network (CNN) Architecture: Employs a convolutional architecture that captures contextual dependencies among spectral features to detect subtle molecular variations across tissue regions.
• Semi-Supervised Approach: Integrates MIL with CNN in a semi-supervised learning strategy to leverage both labeled and unlabeled MSI data and improve classification performance.

Scientific Applications:

• Subtissue classification in MSI: Assigns class labels to subtissue locations within MSI datasets to support molecular spatial mapping.
• Differentiation of tissue types and disease states: Facilitates discrimination between tissue types and disease states based on molecular composition.
• Diagnostic biomarker discovery: Supports identification of spatially localized molecular signatures relevant for diagnostics.
• Pathogenesis studies and targeted therapy development: Aids investigation of pathological mechanisms and development of targeted therapies by resolving subtissue molecular heterogeneity.

Methodology:

Trains a convolutional neural network using multiple instance learning principles on MSI spectra; processes input MSI spectra and assigns class labels to subtissue locations using tissue-level annotations as weak supervision while leveraging labeled and unlabeled data.