Summary
Researchers at OmiCore have released an openly licensed image-analysis model that marks tumour-enriched regions in colorectal cancer tissue slides. In a preprint, the team reports agreement with transcriptomics on one section and describes limits including its 112 μm resolution and lack of validation across scanners and institutions.
OmiCore researchers have released OmiCoreTumorDetector, an openly licensed model that maps tumour-enriched regions in haematoxylin and eosin (H&E) slides of colorectal cancer. In a bioRxiv preprint, they report that the model’s region calls agreed with transcriptome-derived tumour-cell identities on one tissue section, while emphasising that it maps areas rather than individual malignant cells.
H&E is a widely used tissue-staining method: the resulting slides let pathologists examine tissue structure under a microscope. Marking tumour regions by hand is a common first step in spatial-omics work, which measures molecular features while retaining information about where they occur in tissue. The researchers developed the model to automate that region-mapping step and export annotations compatible with the QuPath image-analysis software.
How the model handles slide variation
The released model, omicore-tumordetector-crc-he-v0.1, combines three convolutional classifiers trained on 100,000 public tissue tiles. At inference—the point when a trained model analyses new images—it applies Macenko stain normalisation, a method for adjusting colour variation caused by differences in tissue staining.
The team’s results point to stain variation as an important challenge. During development, a single model achieved an area under the receiver operating characteristic curve (AUROC) of 0.955 on unseen slides, but its sensitivity at the conventional 0.5 decision threshold was 0.48. AUROC summarises how well a model ranks positive cases above negative ones across thresholds; sensitivity at one threshold measures the fraction of positive cases detected at that setting. The contrast illustrates why a strong ranking score alone may not determine how reliably a model marks regions using a fixed cutoff.
On an independently collected tile set, training with non-normalised tiles raised tumour AUROC from 0.836 to 0.992. Applying stain normalisation during inference reduced false-positive tumour area in normal-adjacent tissue by 16- to 26-fold, the authors report.
Tests against tissue slides and molecular data
The released model was tested on five colorectal cancer sections generated with 10x Visium HD, none of which shared material with the training data. It marked 27.7–48.5% of tissue as tumour in three carcinoma sections. In two normal-adjacent sections, it called 0.08% and 1.82% of tissue tumour, respectively, but exported no tumour region from either section.
For one carcinoma section with matched single-cell-resolution transcriptomics, the model’s region scores agreed with transcriptome-derived tumour-cell identities at an AUROC of 0.985. The reported 95% spatial-block bootstrap confidence interval was 0.975–0.993. Because the image model does not use gene-expression data, this comparison provides a separate molecular check on its image-based calls.
The model localises tumour-enriched regions at 112 μm resolution; it is not designed to identify individual malignant cells. The preprint says it has not yet been validated across scanners, institutions or histological variants. OmiCore says the code and weights are available under the Apache-2.0 licence and can be installed with pip install omicoretumordetector. The authors disclose that the work was funded by OmiCore, which developed the model, and that both authors are affiliated with the company. They state that the model is not a medical device.