An effective microscopic image augmentation approach | JoVE Visualize

Area of Science:

Microscopy and Computational Biology
Artificial Intelligence in Botany
Pharmacognosy and Machine Learning

Background:

Chinese medicinal herb (CMH) identification faces challenges due to a vast number of species and difficulties in collecting microscopic images, leading to small sample sizes.
The scarcity of certain cellular features (as low as 0.5%) hinders the effectiveness of deep learning and few-shot learning models.
Expanding data for rare features is crucial for improving CMH identification accuracy.

Purpose of the Study:

To propose an effective microscopic image augmentation approach for few-shot learning (MIAA-FSL) to address data scarcity and class imbalance in CMH identification.
To develop a conditionally guided microscopic image generation model (CGMIGM) for generating rare features.
To integrate semi-supervised learning for data augmentation (SSLDAM) to improve the usability of damaged or blurry microscopic images.

Main Methods:

Conditional guidance using denoising diffusion probabilistic models (DDPM) to generate rare features and mitigate class imbalance.
Semi-supervised learning and pseudo-label generation to enhance and utilize damaged, blurry, or difficult-to-discern microscopic images.
Development of the Microscopic Image Augmentation for Few-Shot Learning (MIAA-FSL) approach combining CGMIGM and SSLDAM.

Main Results:

The MIAA-FSL approach demonstrated an average improvement of 24% in identification accuracy compared to the MIR+DDPM method.
Accuracy in identifying rare features significantly increased from 45.5% to 87.0%.
Effectively mitigated the challenges of object detection with few samples in CMH microscopic image analysis.

Conclusions:

The proposed MIAA-FSL approach effectively addresses the problem of small sample sizes and class imbalance in CMH microscopic image identification.
The combination of conditional generative models and semi-supervised learning enhances the recognition of rare features and improves overall model performance.
This method offers a viable solution for accurate CMH identification, particularly in scenarios with limited data and scarce features.