Deep Model Based Transfer and Multi-Task Learning for Biological Image Analysis

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Developing effective image representations is crucial for learning from image data.
• Traditional texture features are suitable for some tasks like developmental stage determination but not for controlled vocabulary annotation.

Purpose of the Study:

• To develop advanced feature extraction methods for generating hierarchical representations of in situ hybridization (ISH) images.
• To improve the annotation of gene expression patterns using ISH images.

Main Methods:

• Utilized deep convolutional neural networks (CNNs) trained on natural image datasets.
• Implemented a partial transfer learning scheme to adapt models to the ISH image domain.
• Employed multi-task learning to fine-tune pre-trained models with labeled ISH images.

Main Results:

• Deep models incorporating transfer and multi-task learning generated superior feature representations.
• These representations significantly outperformed existing methods for annotating gene expression patterns across different developmental stages.
• The developed methods enable more accurate and detailed annotation of ISH images.

Conclusions:

• Transfer and multi-task learning with deep CNNs provide powerful tools for ISH image analysis.
• The proposed hierarchical representations enhance the ability to annotate gene expression patterns.
• This approach offers a significant advancement in computational approaches to developmental biology research.