Patherea: Cell detection and classification for the 2020s.

Area of Science:

• Computational pathology
• Digital pathology
• Biomedical image analysis

Background:

• Accurate cell detection and classification are crucial for histopathology analysis.
• Existing point-based methods often require intermediate representations and struggle with complex datasets.

Purpose of the Study:

• To introduce Patherea, a unified framework for direct point-based cell detection and classification in histopathology.
• To develop and release the largest manually labeled Ki-67 IHC dataset for method development and benchmarking.
• To improve the accuracy and reproducibility of cell detection evaluations.

Main Methods:

• Patherea employs a unified framework for direct cell location and class prediction without intermediate steps.
• A hybrid Hungarian matching strategy is utilized for precise point assignment.
• The framework supports flexible backbones and leverages pathology foundation models.
• A new, large-scale Ki-67 IHC dataset was created using expert annotations on whole-slide images.

Main Results:

• Patherea achieved superior F1 scores on public datasets (Lizard, BRCA-M2C, BCData) compared to existing point-based methods.
• Performance saturation was observed on current benchmarks, highlighting the need for more challenging datasets.
• The Patherea dataset captures clinically relevant low-abundance cell classes where current methods underperform.
• Common errors in evaluation protocols were identified and corrected, with a benchmarking utility provided.

Conclusions:

• Patherea provides a robust and accurate framework for point-based cell detection and classification in histopathology.
• The new dataset and corrected evaluation protocols facilitate future research and fair comparisons in digital pathology.
• The public release of the dataset and code will advance the field of computational pathology.