Image-Level Uncertainty in Pseudo-Label Selection for Semi-Supervised Segmentation.

Area of Science:

• Deep learning
• Computational pathology
• Medical image analysis

Background:

• Deep learning excels in biomedical image segmentation but requires extensive labeled data.
• Digital pathology presents a challenge due to the abundance of unlabeled data and the cost of expert annotation.
• Semi-supervised learning, particularly pseudo-labeling, offers a way to leverage unlabeled data for improved model training.

Purpose of the Study:

• To adapt image classification pseudo-labeling methods for segmentation tasks in digital pathology.
• To investigate the use of image-level confidence and uncertainty thresholds for selecting pseudo-labels.
• To analyze the relationship between these confidence/uncertainty metrics and model performance.

Main Methods:

• Developed and applied an adapted pseudo-labeling strategy to three digital pathology datasets.
• Explored various image-level confidence and uncertainty thresholds for pseudo-label selection.
• Evaluated the correlation between image-level certainty metrics and segmentation model performance.

Main Results:

• Confidence and uncertainty metrics did not consistently correlate with performance in an intuitive manner.
• Unusual correlations between certainty metrics and performance were identified as indicators of useful pseudo-label generation.
• The adapted approach demonstrated potential for enhancing segmentation model training using unlabeled pathology data.

Conclusions:

• The proposed method effectively adapts image-level confidence and uncertainty for segmentation pseudo-labeling in digital pathology.
• Abnormal correlations in certainty metrics can signal the generation of valuable pseudo-labels.
• Improved segmentation performance facilitates more accurate disease quantification in histopathology.