Image generation by GAN and style transfer for agar plate image segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Deep learning models, particularly Convolutional Neural Networks (CNNs), excel in computer vision tasks like medical image analysis.
• Training CNNs typically requires large annotated datasets, which are scarce and costly in the medical domain.
• Semantic segmentation of bacterial colonies is crucial for infection detection and quantification in Petri plate analysis.

Purpose of the Study:

• To develop a novel approach for semantic segmentation of bacterial colonies in agar plate images.
• To address the challenge of limited annotated medical data by employing synthetic image generation.
• To enhance the training dataset size for deep learning models used in bacterial colony analysis.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) for distinguishing bacterial colonies from the background.
• Developed a generative adversarial network (GAN) to create synthetic bacterial colony images.
• Integrated synthetic data by superimposing bacterial colony patches onto background images, considering visual realism through style transfer algorithms.

Main Results:

• The proposed deep learning method was evaluated on a public dataset for bacterial colony semantic segmentation.
• Incorporating synthetic data into the CNN training yielded performance comparable to using only real images.
• Qualitative results demonstrated the effectiveness of the approach on a second public dataset lacking annotations.

Conclusions:

• Combining a small set of real images with synthetic data achieves results similar to using a full set of real images.
• The synthetic data generator effectively overcomes the limitations of scarce biomedical data.
• This method offers a scalable and economical alternative to manual ground-truth annotation for CNN training.