DEPICTER: Deep representation clustering for histology annotation.

Area of Science:

• Digital pathology
• Computational imaging
• Artificial intelligence in medicine

Background:

• Accurate segmentation of histopathology whole-slide images (WSI) is crucial for cancer diagnosis.
• Fully supervised deep learning methods require extensive manual annotation, which is time-consuming and costly.
• Existing non-fully supervised methods often prioritize algorithmic performance over practical clinical utility.

Purpose of the Study:

• To develop DEPICTER, an interactive tool for efficient and reliable WSI segmentation.
• To leverage self- and semi-supervised learning to reduce the annotation burden for pathologists.
• To provide a practical solution for histopathology image analysis in real-world scenarios.

Main Methods:

• DEPICTER computes patch embeddings using a pretrained model.
• Users interactively select benign and cancerous patches.
• Label propagation is achieved via seeded iterative clustering or feature space gating within the embedding space.

Main Results:

• DEPICTER enables real-time interaction with pathologists.
• The tool produces patch-wise dense segmentation maps at the WSI level.
• Simulations on public datasets demonstrate competitive performance.

Conclusions:

• DEPICTER offers an effective interactive approach for WSI segmentation.
• The tool significantly reduces the workload associated with manual annotation.
• DEPICTER integrates deep representations with user interaction for reliable histopathology analysis.