ClusterSeg

ClusterSeg performs instance segmentation of individual nuclei in densely clustered microscope, cytopathology, and histopathology images to enable accurate delineation of nuclei in challenging biological imaging contexts.

Key Features:

• Hybrid Encoder Architecture: A convolutional-transformer hybrid encoder combines CNN spatial feature extraction with transformers for long-range dependency modeling.
• 2.5-Path Decoder: A 2.5-path decoder predicts nuclei instance masks, contours, and clustered edges to delineate individual nuclei within clusters.
• Clustered-Edge Pointed Annotation Strategy: A clustered-edge pointed annotation strategy targets salient and error-prone boundaries to reduce annotation requirements for training.
• Partially-Supervised PS-ClusterSeg Variant: PS-ClusterSeg leverages the segmentation backbone to improve performance under limited supervision.

Scientific Applications:

• Microscope imaging: Applied to microscope images for nucleus instance segmentation in densely packed cellular scenes.
• Cytopathology imaging: Evaluated on cytopathology images to resolve clustered nuclei in diagnostic preparations.
• Histopathology imaging: Applied to histopathology images to delineate nuclei within tissue sections exhibiting severe clustering.
• Dataset evaluation: Validated using four privately curated image sets and two public datasets characterized by severely clustered nuclei.
• Performance benchmarking: Empirically demonstrates improved accuracy compared with prior state-of-the-art approaches across multiple metrics and modalities.

Methodology:

Integration of a convolutional-transformer hybrid encoder, a 2.5-path decoder, and a clustered-edge pointed annotation strategy, with a partially-supervised PS-ClusterSeg variant that leverages the segmentation backbone for limited supervision.