STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology.

Area of Science:

• Digital Pathology
• Machine Learning in Oncology
• Computational Biology

Background:

• Accurate colorectal cancer (CRC) tissue classification is crucial for machine learning-driven diagnosis and treatment.
• Existing CRC datasets lack morphologic diversity, class balance, and image quality, hindering model performance and generalizability.

Purpose of the Study:

• Introduce STARC-9, a large-scale dataset for multi-class CRC tissue classification.
• Address limitations of current datasets by providing diverse, high-quality histopathologic image tiles.
• Validate the utility of STARC-9 for improving downstream machine learning models.

Main Methods:

• Developed DeepCluster++, a novel framework for semi-automated dataset curation.
• Utilized autoencoder-based feature extraction and K-means clustering for tile grouping.
• Employed equal-frequency binning and expert pathologist verification to ensure tile diversity and accuracy.

Main Results:

• Created STARC-9 with 630,000 histopathologic tiles across nine CRC tissue classes.
• Demonstrated superior generalizability of machine learning models trained on STARC-9 compared to those trained on existing datasets.
• Validated STARC-9's effectiveness on multi-class CRC tissue classification and segmentation tasks.

Conclusions:

• STARC-9 significantly enhances the performance and generalizability of machine learning models for CRC tissue classification.
• The DeepCluster++ framework offers a flexible approach for constructing high-quality datasets from whole-slide images (WSI) across various applications.
• STARC-9 facilitates advancements in AI-powered cancer diagnostics and personalized medicine.