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

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Colorectal cancer (CRC) histopathology image classification is crucial for machine learning-driven diagnosis and treatment.
• Existing CRC datasets often lack diversity, exhibit class imbalance, and contain low-quality images, hindering model performance and generalizability.

Purpose of the Study:

• Introduce STARC-9, a large-scale dataset for multi-class tissue classification in CRC.
• Address limitations of current datasets by providing morphologically diverse, balanced, and high-quality image tiles.

Main Methods:

• Developed STARC-9 with 630,000 H&E-stained image tiles across nine CRC tissue classes (70,000/class) from 200 patients.
• Utilized DeepCluster++ framework: histopathology autoencoder for feature extraction, K-means clustering for tile grouping, and equal-frequency binning for diversity.
• Incorporated expert pathologist verification for selected tiles, ensuring accuracy and reducing manual curation.

Main Results:

• Benchmarked deep learning models (CNNs, transformers, foundation models) on STARC-9 for classification and segmentation tasks.
• Models trained on STARC-9 demonstrated superior generalizability compared to those trained on existing datasets.
• Validated the DeepCluster++ framework's effectiveness in creating high-quality, diverse datasets.

Conclusions:

• STARC-9 provides a valuable resource for advancing CRC histopathology image analysis and machine learning applications.
• The DeepCluster++ framework offers a flexible and efficient method for constructing large-scale, high-quality datasets from whole-slide images.
• This approach has broad applicability beyond CRC for various medical imaging datasets.