OmiCoreTumorDetector: an open, molecularly validated model for mapping tumour regions in colorectal cancer H&E sections
- 1Omicore, Inc
Abstract
Defining tumour regions on haematoxylin and eosin (H&E) sections is a routine first step in spatial-omics studies, yet it is usually done by hand and is difficult to reproduce. We present OmiCoreTumorDetector, an openly licensed model that maps tumour-enriched regions in colorectal cancer (CRC) H&E sections and exports them as QuPath-compatible annotations. The released model (omicore-tumordetector-crc-he-v0.1) is an ensemble of three convolutional classifiers trained on 100,000 public tissue tiles, combined with Macenko stain normalisation at inference. During development we found that the main obstacle to reuse was calibration under stain-domain shift rather than discrimination: a single model kept an area under the ROC curve (AUROC) of 0.955 on unseen slides while its sensitivity at the conventional 0.5 threshold fell to 0.48. Training on non-normalised tiles raised tumour AUROC on an independently collected tile set from 0.836 to 0.992, and normalising at inference reduced false-positive tumour area in normal-adjacent tissue by 16- to 26-fold. On five 10x Visium HD CRC sections that share no material with the training data, the released model called 27.7-48.5% of tissue as tumour in three carcinoma sections and 0.08% and 1.82% in two normal-adjacent sections, exporting no tumour region from either normal section. On the carcinoma section with matched single-cell-resolution transcriptomics, agreement with transcriptome-derived tumour-cell identities reached an AUROC of 0.985 (95% spatial-block bootstrap CI 0.975-0.993). The image model never observes gene expression, so this is orthogonal evidence. The model localises tumour-enriched regions at 112 um resolution; it does not identify individual malignant cells and has not yet been validated across scanners, institutions or histological variants. Code, weights and evaluation are released under Apache-2.0 and installable with pip install omicoretumordetector.
Related
- Biopsy-Validated Deep Learning Mapping of Glioma Infiltration Beyond Contrast Enhancement Using Multiparametric MRI
- Improving Micro– and Macro–Algae Speciation Mapping for Management: A Field‐Validated Sentinel‐2 and Sentinel‐3 Intercomparison in the Indian River Lagoon
- Histologically Validated Ex Vivo Characterization of Intraplaque Hemorrhage and Lipid Components Using Quantitative Susceptibility Mapping
- A Spatially Validated Machine-Learning Framework for Mapping Alpha-Gal Syndrome Exposure Potential in Illinois
- 1132 Spatial mapping and significance of the IL-12 pathway in human colorectal cancer
BibTeX
@misc{niwase2026omicoretumordetector,
title = {{OmiCoreTumorDetector: an open, molecularly validated model for mapping tumour regions in colorectal cancer H&E sections}},
author = {S. Niwase and A. Fujiyama},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.21.753083},
url = {https://www.biorxiv.org/content/10.64898/2026.09.21.753083v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - OmiCoreTumorDetector: an open, molecularly validated model for mapping tumour regions in colorectal cancer H&E sections
AU - S. Niwase
AU - A. Fujiyama
PY - 2026
DA - 2026/09/21
PB - bioRxiv
DO - 10.64898/2026.09.21.753083
UR - https://www.biorxiv.org/content/10.64898/2026.09.21.753083v1
AB - Defining tumour regions on haematoxylin and eosin (H&E) sections is a routine first step in spatial-omics studies, yet it is usually done by hand and is difficult to reproduce. We present OmiCoreTumorDetector, an openly licensed model that maps tumour-enriched regions in colorectal cancer (CRC) H&E sections and exports them as QuPath-compatible annotations. The released model (omicore-tumordetector-crc-he-v0.1) is an ensemble of three convolutional classifiers trained on 100,000 public tissue tiles, combined with Macenko stain normalisation at inference. During development we found that the main obstacle to reuse was calibration under stain-domain shift rather than discrimination: a single model kept an area under the ROC curve (AUROC) of 0.955 on unseen slides while its sensitivity at the conventional 0.5 threshold fell to 0.48. Training on non-normalised tiles raised tumour AUROC on an independently collected tile set from 0.836 to 0.992, and normalising at inference reduced false-positive tumour area in normal-adjacent tissue by 16- to 26-fold. On five 10x Visium HD CRC sections that share no material with the training data, the released model called 27.7-48.5% of tissue as tumour in three carcinoma sections and 0.08% and 1.82% in two normal-adjacent sections, exporting no tumour region from either normal section. On the carcinoma section with matched single-cell-resolution transcriptomics, agreement with transcriptome-derived tumour-cell identities reached an AUROC of 0.985 (95% spatial-block bootstrap CI 0.975-0.993). The image model never observes gene expression, so this is orthogonal evidence. The model localises tumour-enriched regions at 112 um resolution; it does not identify individual malignant cells and has not yet been validated across scanners, institutions or histological variants. Code, weights and evaluation are released under Apache-2.0 and installable with pip install omicoretumordetector.
ER -