CellViT: Vision Transformers for precise cell segmentation and classification.

Area of Science:

• Computational pathology
• Digital pathology
• Biomedical image analysis

Background:

• Nuclei detection and segmentation in H&E stained tissue images are critical for clinical applications.
• Challenges include variations in nuclear staining, size, overlapping boundaries, and clustering.
• Convolutional neural networks (CNNs) are widely used, but Transformer-based networks offer new potential.

Purpose of the Study:

• To introduce CellViT, a novel deep learning method for automated instance segmentation of cell nuclei.
• To explore the efficacy of Transformer-based networks with large-scale pre-training for nuclei segmentation.
• To achieve state-of-the-art performance on challenging nuclei instance segmentation datasets.

Main Methods:

• Developed CellViT, a deep learning architecture based on Vision Transformer for nuclei instance segmentation.
• Trained and evaluated CellViT on the PanNuke dataset, a large and complex dataset with ~200,000 annotated nuclei.
• Leveraged large-scale pre-trained Vision Transformers, including the Segment Anything Model and a ViT-encoder trained on 104 million histological image patches.

Main Results:

• Achieved state-of-the-art nuclei detection and instance segmentation performance on the PanNuke dataset.
• Obtained a mean panoptic quality of 0.50 and an F1-detection score of 0.83.
• Demonstrated the superiority of large-scale in-domain and out-of-domain pre-trained Vision Transformers.

Conclusions:

• CellViT, utilizing pre-trained Vision Transformers, significantly advances automated cell nuclei instance segmentation.
• The method addresses key challenges in nuclei segmentation, offering improved accuracy in digital pathology.
• The publicly available code facilitates further research and application in computational pathology.