TOP-GAN: Stain-free cancer cell classification using deep learning with a small training set

Area of Science:

• Medical Imaging
• Deep Learning
• Biotechnology

Background:

• Deep learning in medical imaging is hindered by small training datasets.
• Accurate cell classification is crucial for cancer diagnosis and research.

Purpose of the Study:

• To develop a deep learning approach that overcomes the challenge of limited training data for medical image classification.
• To apply this method for accurate classification of healthy and cancerous cell lines using quantitative phase imaging.

Main Methods:

• Proposed a novel method named transferring of pre-trained generative adversarial network (TOP-GAN), a hybrid of transfer learning and generative adversarial networks (GANs).
• Utilized GANs to train on a large dataset of unclassified sperm cells to address the scarcity of classified medical images.
• Extracted optical path delay maps from low-coherence off-axis holography images of cells as network inputs.
• Modified the final network layers to create automatic classifiers for healthy, primary cancer, and metastatic cancer cells.

Main Results:

• Achieved high classification accuracies of 90-99% for cell types, even with training sets as small as several images.
• Demonstrated superior performance compared to traditional methods designed for small training set problems.
• Enabled rapid, automatic, and accurate classification in stain-free imaging flow cytometry.

Conclusions:

• The TOP-GAN approach makes holographic microscopy and deep learning more accessible for medical applications.
• The method shows significant potential for various medical image classification tasks facing limited data availability.
• This technique facilitates accurate cell classification without the need for staining.