Deep Low-Shot Learning for Biological Image Classification and Visualization From Limited Training Samples.

Area of Science:

• Computational Biology
• Developmental Biology
• Machine Learning

Background:

• Biological image analysis, particularly for gene expression patterns, requires precise developmental stage classification.
• High costs and time associated with labeling training data limit the development of accurate predictive models.
• Identifying developmental landmarks is crucial for interpreting model predictions and refining biological insights.

Purpose of the Study:

• To develop an accurate computational model for precise developmental stage classification of in situ hybridization (ISH) images using limited training data.
• To address the challenge of building predictive models when labeled biological image datasets are scarce.
• To enable the identification and visualization of developmental landmarks for biological interpretation.

Main Methods:

• Proposed a deep two-step low-shot learning framework for ISH image classification.
• Implemented a novel approach involving both data-level and feature-level learning.
• Utilized a deep residual network as the base model and employed saliency maps for interpretation.

Main Results:

• Achieved improved performance in precise developmental stage prediction for ISH images.
• Demonstrated the model's ability to provide biologically meaningful interpretations through saliency maps.
• Validated the model's accuracy and interpretability in classifying limited training samples.

Conclusions:

• The proposed deep low-shot learning framework effectively classifies ISH images with limited training data.
• Saliency maps derived from the model aid in identifying and visualizing key developmental landmarks.
• The methodology shows promise for generalization to other biological image classification tasks with small datasets.