Generic Isolated Cell Image Generator.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automated cancer screening systems using image analysis are crucial but hindered by limited labeled medical data.
• Deep learning models require extensive datasets, which are scarce in medicine due to privacy and labeling challenges.
• Existing data augmentation techniques offer limited diversity, necessitating advanced generative methods.

Purpose of the Study:

• To develop a novel, generic generator for 2D cell images to address data scarcity in automated bladder cancer screening.
• To enhance the training dataset for deep learning models used in analyzing urinary cytology digital slides.
• To explore the application of doodle-style transfer for synthetic medical image generation.

Main Methods:

• Developed a new framework for 2D cell image generation by combining existing techniques with a novel doodle-style transfer method.
• Applied the framework to synthesize cell images for bladder cancer screening applications.
• Conducted statistical evaluations to compare feature distributions of real and synthetic images.
• Performed visual assessments with medical experts to evaluate the realism of generated images.

Main Results:

• The developed framework successfully generated synthetic 2D cell images.
• Statistical analysis indicated that features of real and synthetic cell images largely followed similar distributions.
• Medical experts visually confirmed the realism of the synthetic cell images.
• The modular framework shows potential for application in other cell image generation tasks.

Conclusions:

• The novel doodle-style transfer-based framework effectively generates realistic synthetic cell images.
• This method addresses the challenge of limited training data for deep learning in medical imaging, specifically for bladder cancer screening.
• The approach offers a promising solution for augmenting medical datasets and improving the performance of automated diagnostic systems.