Novel transfer learning schemes based on Siamese networks and synthetic data

Area of Science:

• Biotechnology
• Computer Vision
• Machine Learning

Background:

• Deep learning transfer learning excels in computer vision but struggles with dissimilar data, like in biotechnology.
• Analyzing Chinese Hamster Ovary (CHO-K1) cell growth in microfluidics presents unique data challenges.
• Limited labeled data and high annotation costs hinder traditional deep learning applications in this domain.

Purpose of the Study:

• To develop a novel transfer learning scheme for automatic analysis of CHO-K1 suspension growth in microfluidic single-cell cultivation.
• To adapt deep network models to biotechnology domains with dissimilar data characteristics.
• To address scenarios with scarce or absent data labels.

Main Methods:

• Proposed a novel transfer learning scheme expanding the Twin-Variational Autoencoder (Twin-VAE) architecture.
• Trained the Twin-VAE on both realistic and synthetic data, modifying the training procedure for transfer learning.
• Investigated a strategy incorporating simultaneous retraining on natural and synthetic data using an invariant shared representation and target variables.

Main Results:

• The modified Twin-VAE architecture demonstrated superiority over state-of-the-art transfer learning and classical image processing methods.
• The approach achieved satisfactory results even with significantly reduced training times.
• The method successfully handled unseen data from different microscopy technologies.

Conclusions:

• The proposed Twin-VAE transfer learning scheme is effective for analyzing CHO-K1 cell growth in microfluidics, outperforming existing methods.
• This approach offers a viable solution for domains with limited labeled data and unique data characteristics.
• The open-source code and available datasets facilitate further research and application.