Semi-supervised tissue segmentation from histopathological images with consistency regularization and uncertainty...

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Pathologist assessment of histopathological smear images is time-consuming and prone to errors.
• Deep learning models like Convolutional Neural Networks (CNNs) can automate tissue structure analysis.
• Training deep learning models requires large labeled datasets, which are often scarce, especially for rare diseases.

Purpose of the Study:

• To develop a novel semi-supervised method for semantic segmentation of tissue structures in histopathological images.
• To address the challenge of limited labeled data in training deep learning models for medical image analysis.
• To improve the accuracy and consistency of automated histopathological image analysis.

Main Methods:

• A CNN-based teacher model was developed to generate pseudo-labels for training a student model.
• Self-supervised training was employed to enhance the teacher model's performance on smaller datasets.
• Consistency regularization was used for efficient student model training on labeled data.
• Monte Carlo dropout was utilized for uncertainty estimation of the model's predictions.

Main Results:

• The proposed semi-supervised method achieved a mean Intersection over Union (mIoU) score of 0.64 on a public dataset.
• The approach demonstrated the potential to overcome limitations of traditional supervised learning in histopathology.
• The model showed promising results in segmenting tissue structures, indicating improved accuracy.

Conclusions:

• The developed semi-supervised learning framework offers a viable solution for histopathological image segmentation with limited data.
• The teacher-student model architecture effectively leverages pseudo-labeling and self-supervision.
• This work highlights the potential of AI to enhance diagnostic accuracy and efficiency in digital pathology.