Research map / Label-efficient and robust learning / Recognition and long-tail
STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces STARC-9, a 630,000-tile, nine-class H&E colorectal cancer tissue dataset built with DeepCluster++ (autoencoder features, K-means, equal-frequency binning, pathologist verification) and benchmarks many models trained on it versus NCT-CRC-HE-100K and HMU-GC-HE-30K.
Problem
Publicly available CRC histopathology tile datasets for multi-class tissue classification suffer from limited morphologic diversity, class imbalance, and low-quality or mislabeled tiles, and manual tile curation is labor-intensive.
Approach
A semi-automated pipeline (DeepCluster++) that clusters SSIM-trained autoencoder embeddings of tiles from each WSI with K-means, then samples across equal-frequency distance bins within each cluster, followed by pathologist verification, to build a balanced, morphologically diverse dataset.
Key results
- CTransPath trained on STARC-9 reached 99.00% accuracy on STANFORD-CRC-HE-VAL-LARGE, versus 84.25% for ViT-base trained on NCT and 91.99% for HiPT trained on HMU.
- STARC-9-trained models achieved mean tumor Dice of 90.47% on the Stanford set and 89.38% on the TCGA-CRC set, about 14% and 17% above NCT and HMU on Stanford.
- On the IMP-CRS10K test set of seven classes, the STARC-9-trained model reached 96.61% accuracy and 95.55% F1-macro.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train the model on STARC-9 instead of NCT | EfficientNet-B7 on CRC tissue classification · STANFORD-CRC-HE-VAL-LARGE | accuracy: 82.47 -> 98.80 | Helped |
| train the model on STARC-9 instead of HMU | EfficientNet-B7 on CRC tissue classification · STANFORD-CRC-HE-VAL-LARGE | accuracy: 84.45 -> 98.80 | Helped |
| train CTransPath on STARC-9 instead of NCT | CTransPath on CRC tissue classification · STANFORD-CRC-HE-VAL-LARGE | accuracy: 79.05 -> 99.00 | Helped |
| train CTransPath on STARC-9 instead of HMU | CTransPath on CRC tissue classification · STANFORD-CRC-HE-VAL-LARGE | accuracy: 91.88 -> 99.00 | Helped |
| train the model on STARC-9 instead of NCT | CTransPath on tumor segmentation · Stanford | Dice score (%): 78.20 ± 17.01 -> 90.47±8.14 | Helped |
| train the model on STARC-9 instead of HMU | CTransPath on tumor segmentation · TCGA-CRC | Dice score (%): 68.85 ± 22.10 -> 89.38 ± 9.14 | Helped |
| use SSIM loss instead of MSE loss (ablation) | AE_CRC · STANFORD-CRC-HE-VAL-SMALL | SSIM: 0.8863 -> 0.9262 | Helped |
| use SSIM loss instead of MSE loss (ablation) | AE_CRC · STANFORD-CRC-HE-VAL-SMALL | pixel error: 0.0015 -> 0.0012 | Helped |
| use SSIM loss instead of MSE loss (ablation) | AE_CRC · STANFORD-CRC-HE-VAL-SMALL | PSNR: 28.53 dB -> 32.48 dB | Helped |
| set samples per cluster to 200 instead of 400 (ablation) | DeepCluster++ · tumor and necrosis cluster quality | Average Entropy: 0.12 at m=200; 0.41 at m=400 | Mixed |
| increase samples per cluster from 400 to 600 (ablation) | DeepCluster++ · tumor and necrosis cluster quality | Average Entropy: 0.41 -> 1.73 | Mixed |
| increase samples per cluster from 400 to 800 (ablation) | DeepCluster++ · tumor and necrosis cluster quality | Average Entropy: 0.41 -> 2.17 | Hurt |
Surprising findings
- Medium-sized transformers around 80-90M parameters performed best, and larger models such as Prov-GigaPath and Virchow (305M) gave no significant gains.
- Off-the-shelf pathology foundation models like UNI were judged to over-separate biologically related tissues, so a custom SSIM-trained autoencoder was used for sampling features instead.
Limitations
- Single-institution data from Stanford with limited Black and Native American representation.
- Limited to CRC and nine tissue classes; may not cover all CRC tissue types or generalize to other cancers.
- Equal-per-class tile count of 70,000 is a design choice that may not match real-world class distributions; cluster-to-class mapping required manual labeling.
Experimental setup
Datasets: STARC-9 NCT-CRC-HE-100K HMU-GC-HE-30K STANFORD-CRC-HE-VAL-SMALL STANFORD-CRC-HE-VAL-LARGE CURATED-TCGA-CRC-HE-VAL-20K IMP-CRS10K TCGA-CRC
Benchmarks: STANFORD-CRC-HE-VAL-LARGE STANFORD-CRC-HE-VAL-SMALL CURATED-TCGA-CRC-HE-VAL-20K IMP-CRS10K Stanford tumor segmentation set TCGA-CRC tumor segmentation set
Metrics: precision recall F1-macro accuracy IoU Dice score number of parameters
Baselines: ResNet50 EfficientNet-B7 KimiaNet ViT-base DeiT-B Swin Transformer-base ConvNeXT-base CTransPath HiPT Prov-Gigapath Path-DINO CONCH
Builds on: K-means clustering NCT-CRC-HE-100K HMU-GC-HE-30K Macenko normalization
Closest papers by meaning
- Rotation-Agnostic Image Representation Learning for Digital Pathology (CVPR 2024 · similarity 0.86)
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.86)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.85)
- Multi-Resolution Pathology-Language Pre-training Model with Text-Guided Visual Representation (CVPR 2025 · similarity 0.84)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.84)
- MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.