Deep multiple instance learning classifies subtissue locations in mass spectrometry images from tissue-level...

Area of Science:

• Computational pathology
• Molecular imaging
• Biomedical data analysis

Background:

• Mass spectrometry imaging (MSI) provides spatial molecular data for tissue analysis.
• Supervised classification of MSI data requires precise subtissue labels, which are difficult and expensive to obtain.
• Existing classifiers trained with limited or approximate labels exhibit suboptimal performance.

Purpose of the Study:

• To develop a novel semi-supervised approach for accurate subtissue classification in MSI data.
• To leverage weak, tissue-level annotations for improving classification performance.
• To enable rapid distinction between tissue types and disease states using MSI.

Main Methods:

• A semi-supervised method named mi-CNN (multiple instance Convolutional Neural Network) was developed.
• mi-CNN utilizes multiple instance learning to enable weak supervision from tissue-level annotations.
• The CNN architecture captures contextual dependencies within spectral features for classification.

Main Results:

• mi-CNN demonstrated improved subtissue classification performance compared to traditional classifiers.
• Evaluations on simulated and experimental datasets validated the effectiveness of the approach.
• The method successfully utilizes weak supervision for enhanced MSI data analysis.

Conclusions:

• mi-CNN represents a significant advancement in MSI subtissue classification.
• The approach alleviates the challenge of limited ground truth labels in MSI datasets.
• Accurate subtissue classification with mi-CNN facilitates rapid tissue and disease state differentiation.