Few-shot Tumor Bud Segmentation Using Generative Model in Colorectal Carcinoma

Area of Science:

• Computational pathology
• Artificial intelligence in histopathology
• Medical image analysis

Background:

• Deep learning in histopathology faces limitations due to small datasets and extensive labeling times.
• Accurate quantification of tumor budding (TB) in colorectal cancer (CRC) is vital for staging and prognosis but is hindered by manual annotation's labor intensity and potential bias.

Purpose of the Study:

• To develop an annotation-efficient method for generating a large-scale dataset for training AI models for tumor budding detection and quantification.
• To overcome the challenges of limited data and manual annotation in current deep learning approaches for CRC histopathology.

Main Methods:

• A DatasetGAN-based approach was employed to generate synthetic H&E-stained colon tissue images with tumor budding masks.
• The generation process utilized a moderate number of unlabeled images and a few annotated images.
• A downstream UNet++ segmentation model was trained using the synthetically generated images and masks.

Main Results:

• The generated images closely mimic real colon tissue H&E-stained slides.
• The UNet++ model trained on generated data achieved reasonable tumor budding segmentation performance, particularly at the instance level.
• The study demonstrated the feasibility of using GANs for data augmentation in histopathology.

Conclusions:

• DatasetGAN can generate a virtually unlimited supply of annotated images for training AI models, addressing data scarcity in histopathology.
• This annotation-efficient approach shows significant potential for developing robust automatic tumor budding detection and quantification systems.
• The findings pave the way for improved AI-driven diagnostic tools in colorectal cancer research and clinical practice.