Generative Modeling of Histology Tissue Reduces Human Annotation Effort for Segmentation Model Development

Area of Science:

• Digital pathology
• Medical image analysis
• Computational biology

Background:

• Accurate segmentation of histology whole slide images is crucial for tissue analysis.
• Manual annotation of training datasets is time-consuming and labor-intensive.
• Semi-supervised learning offers a promising approach to reduce annotation burden.

Purpose of the Study:

• To evaluate the efficacy of datasetGAN, a semi-supervised learning method, for glomeruli segmentation in renal biopsy images.
• To compare the annotation requirements and performance of datasetGAN against traditional supervised methods.
• To assess the utility of datasetGAN for transfer learning in histology image segmentation.

Main Methods:

• Application of the datasetGAN semi-supervised learning framework for image segmentation.
• Training segmentation models using both labeled and unlabeled renal biopsy image data.
• Comparative analysis of model performance based on varying amounts of annotated data.
• Evaluation of datasetGAN's effectiveness in transfer learning scenarios.

Main Results:

• DatasetGAN significantly reduces the amount of annotated data required for high-performing glomeruli segmentation.
• Models trained with datasetGAN achieve comparable or superior performance to those trained with extensive traditional annotation.
• DatasetGAN substantially improves segmentation model performance when limited whole slide images are available for training, indicating strong transfer learning capabilities.

Conclusions:

• DatasetGAN is an effective semi-supervised method for reducing annotation effort in histology image segmentation.
• This approach holds significant potential for accelerating tissue analysis and improving the efficiency of developing diagnostic tools.
• DatasetGAN shows promise for enhancing transfer learning, enabling robust model training even with scarce annotated datasets.