DEPICTER

DEPICTER performs interactive, patch-wise segmentation and seeded label propagation of whole-slide histopathology images (WSIs) using pretrained embeddings and self- and semi-supervised learning to produce dense WSI-level segmentation maps that distinguish benign and cancerous tissue.

Key Features:

• Interactive Segmentation: Enables pathologist-guided, iterative segmentation of WSIs through selection of representative patches at multiple resolutions.
• Patch-wise Dense Segmentation Map: Generates a dense segmentation map at the WSI level by analyzing image patches to classify tissue regions as benign or cancerous.
• Pretrained Model for Embeddings: Uses a pretrained model to compute deep embeddings from image patches that serve as feature representations for downstream analysis.
• User-Driven Label Propagation: Implements seeded iterative clustering and feature space gating to propagate labels from user-selected patches across the embedding space.
• Self- and Semi-supervised Learning: Integrates self- and semi-supervised learning approaches to reduce reliance on exhaustive pixel-level annotations.
• Evaluation: Has been evaluated in real-time interactions with three pathologists and via simulations on three public cancer classification dataset benchmarks.

Scientific Applications:

• Histopathology Segmentation and Diagnosis: Supports tissue segmentation for diagnostic analysis by distinguishing benign and cancerous regions in WSIs.
• Annotation-efficient Dataset Creation: Reduces the need for extensive pixel-level annotations, enabling scalable annotation of large WSI cohorts.
• Research and Clinical Studies: Facilitates large-scale histopathology studies and the integration of pathologist expertise with machine learning in clinical and research contexts.

Methodology:

A pretrained model computes embeddings from image patches; users select representative benign and cancerous patches at multiple resolutions; labels are propagated across the embedding space using seeded iterative clustering or feature space gating to produce a patch-wise dense WSI segmentation, with self- and semi-supervised learning approaches employed to limit pixel-level annotation requirements.