Classifying and segmenting microscopy images with deep multiple instance learning.

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning

Background:

• High-content screening (HCS) generates vast microscopy datasets requiring automated analysis.
• Existing deep learning models are not optimized for complex microscopy images with multiple objects.
• Automated image analysis is crucial for extracting meaningful data from HCS experiments.

Purpose of the Study:

• To develop a novel deep learning approach for classifying and segmenting microscopy images.
• To enable accurate analysis of cell populations using only whole-image annotations.
• To overcome limitations of existing models in handling complex biological imaging data.

Main Methods:

• A novel neural network architecture combining Convolutional Neural Networks (CNNs) with Multiple Instance Learning (MIL).
• Introduction of the Noisy-AND pooling function for robust aggregation of CNN feature maps.
• End-to-end training of MIL-CNNs using only image-level labels, eliminating the need for manual segmentation.

Main Results:

• The proposed MIL-CNN approach successfully classifies and segments microscopy images containing cell populations.
• The method demonstrates superior performance compared to previous approaches on both mammalian and yeast datasets.
• The Noisy-AND pooling function proved robust to outliers during feature aggregation.

Conclusions:

• The integrated CNN-MIL framework provides an effective solution for analyzing large-scale microscopy data.
• This method advances automated image analysis in cell biology and drug screening.
• The approach simplifies the analysis pipeline by removing the requirement for pre-segmentation.