Can AI predict five-year colorectal cancer survival from pathology slides?
A retrospective preprint reports moderate discrimination across a US trial cohort and an external NHS dataset. The model is not ready to guide treatment, and its selected ensemble did not clearly beat the strongest comparator.
The short version
3 key points
- 1.The study used 562 TCGA cases to examine tissue types, 424 stage III colon-cancer cases from a US clinical trial for development and 218 NHS Lothian cases for external evaluation.
- 2.The single model reached an AUC of 0.642 in internal cross-validation and 0.669 externally; a post hoc ensemble reached 0.70 internally and 0.678 externally, where its interval overlapped the strongest baseline.
- 3.This was retrospective, unreviewed and complete-case research. It excluded early-censored and mucinous cases from development and did not test treatment decisions, prospective outcomes or benefit beyond current clinical practice.
The answer: the slides carried a signal, but not enough for clinical use
An AI system trained on digitised colorectal-cancer slides separated people with different five-year outcomes better than chance in two retrospective cohorts. Its external AUC was about 0.67, meaning it ranked a randomly selected survivor above a randomly selected non-survivor roughly two times in three. That is a research signal, not a treatment rule.
The strongest reported result came from an ensemble selected after reviewing out-of-fold performance. It reached an internal AUC of 0.70 and an external AUC of 0.678. On the external NHS Lothian cohort, that point estimate was essentially the same as the best comparison method, FR-MIL at 0.6766, and their confidence intervals overlapped.
For patients and oncologists, the practical conclusion is restrained. Routine slides may contain prognostic information outside the tumour epithelium, including stroma and adipose tissue. The study does not show that adding this score changes chemotherapy choices, improves survival or works safely across the populations seen in everyday practice.[1]
Three cohorts answered different questions
The researchers used 562 TCGA colon and rectal cancer cases only to ask which tissue morphologies carried predictive signal. Models limited to adipose tissue reached an AUC of 0.622, stroma 0.604 and tumour epithelium about 0.60. Those are weak individual signals, but their spread across tissue types motivated a model that learns from the whole slide.
Development used 424 people from the Alliance/CALGB 89803 phase III trial in the United States. Every person had completely resected stage III colon cancer, and 103 died within five years while 321 survived at least five years. A pathologist blinded to outcomes confirmed tumour was present, and slides underwent checks for staining, focus and scanning artefacts.
External evaluation used 218 colon-cancer cases from the public SurGen resource collected through NHS Lothian. The model was trained only on Alliance data and applied without fine-tuning. The different scanner, resolution, staining and case mix provide a meaningful distribution shift, although one external cohort cannot establish broad clinical generalisability.[1]
How MorphContrast was trained
A whole-slide image contains thousands of small patches but only one patient-level outcome. MorphContrast first uses a frozen pathology encoder to label patch morphology. It then learns to place patches together or apart according to tissue type and whether patients share the same five-year outcome, while an attention layer combines patch information into one risk score.
Within a patient, matching tissue types act as positive pairs and different morphologies as negatives. Across patients, morphology-matched patches with the same outcome are positives, while matched morphologies with different outcomes become hard negatives. A diversity term is intended to prevent the representation collapsing onto one feature.
The primary analysis used five patient-disjoint folds. Hyperparameters, loss functions, ensemble sizes, aggregation rules and thresholds were explored. The selected three-member ensemble was chosen from 50 candidates using an objective that rewarded AUC and balanced accuracy while penalising an imbalance between sensitivity and specificity.[1]
What the reported performance means
The single MorphContrast model averaged an AUC of 0.642 across the internal folds, with 63.5% sensitivity and 55.4% specificity. The selected ensemble produced a pooled out-of-fold AUC of 0.70, with a 95% bootstrap interval from 0.65 to 0.76, and balanced sensitivity and specificity of 68%.
A hazard ratio of 3.79 separated the ensemble's high- and low-risk groups, but that threshold was selected within the development process. Hazard-ratio size can also look impressive when groups are uneven: another comparator produced a higher ratio while assigning only nine people to its high-risk group. Discrimination, calibration and decision benefit all need to be judged together.
On the external set, the single model reached an AUC of 0.669 and the ensemble 0.678. The ensemble's sensitivity was 65.2% and specificity 66.5%, with wide confidence intervals. About one-third of survivors and one-third of non-survivors would therefore fall on the wrong side of that operating point if it were used as a binary classifier.[1]
The comparison is useful but does not prove added clinical value
MorphContrast outperformed ten frozen pathology encoders using the same attention head, whose mean AUCs ranged from 0.51 to 0.63. Yet the single model underperformed two morphology-aware baselines: MorphDistill reached 0.68 and PRISM 0.70. Only the post hoc ensemble matched PRISM internally.
Clinical-variable logistic models built from the available trial data reached pooled AUCs near 0.54 to 0.55. That comparison does not establish superiority over a modern clinical prognostic workflow. The preprint does not compare against current multidisciplinary assessment, molecular testing or validated calculators using all information available to an oncologist.
A model can improve statistical ranking without improving a decision. Before adoption, researchers would need to specify the clinical action triggered by a score, the harms of false high- and low-risk classifications, whether performance is calibrated across groups, and whether using the score improves outcomes in a prospective trial.[1]
Who was excluded and why that matters
The development cohort excluded people lost to follow-up before five years and those censored earlier, turning survival into a complete-case binary label. That discards time-to-event information and can introduce selection bias. It also excluded 50 mucinous cancers because their architecture complicated one morphological assessment.
The trial cohort was narrow: all participants had stage III colon cancer and were fit enough for an adjuvant chemotherapy trial. The paper's supplementary table reports 391 White and 33 Black participants. Performance across other racial groups, health systems, cancer stages and rectal cancers is therefore unknown.
The external subset broadens scanner and case-mix variation but still contains only 218 colon-cancer cases. A single archival slide represented more than 98% of development cases, which may not capture within-tumour heterogeneity or the range of slide selection practices across hospitals.[1]
What clinicians and patients should take from it
The study supports looking beyond malignant cells. Signals from stroma, adipose tissue and other surrounding structures may add prognostic information, and morphology-aware training offers a way to organise it. That is a promising research direction for digital pathology departments already scanning routine slides.
It should not yet change an individual's prognosis or treatment. The model's errors remain substantial, code is promised only after acceptance, and the paper has not undergone peer review. A risk score also needs understandable uncertainty and a route for a pathologist to inspect which regions influenced it.
The work was supported by the US National Cancer Institute, Pelotonia and Ohio State institutions. The authors declared no financial or non-financial competing interests. The trial slides are available only to qualified investigators on request, while TCGA and SurGen data are public.[1]
What would change the assessment
Confidence would increase with locked-model validation across multiple hospitals, scanners and demographic groups, including mucinous disease and patients with incomplete follow-up. Calibration, subgroup error rates and comparison with a full contemporary clinical baseline should be reported before claims of added value.
The decisive evidence would be a prospective study showing how clinicians use the score and whether it improves a specified decision without worsening inequity or overtreatment. A randomised impact trial would then need to show benefit for patients, not only a higher AUC.
Until that evidence exists, the fairest conclusion is that routine pathology slides contain a reproducible but moderate survival signal. Morphology-aware learning may help extract it, but the reported model remains a research instrument rather than a clinical prognostic test.[1]
What this means for people
- Patients should not interpret the reported score as an individual prognosis; the model was retrospective and made substantial errors at its reported operating point.
- Pathologists may gain a research method for using information across tumour and surrounding tissue, but prospective workflow and outcome studies are still missing.
- The narrow trial cohort and small external sample leave performance across hospitals, stages and demographic groups uncertain.
Global context
The model was developed on a US stage III clinical-trial cohort and externally tested on an NHS Lothian dataset. That cross-system test is useful, but broader validation is needed across countries, scanners, stages, tumour subtypes and patient groups.
What the evidence does not yet show
- The paper is an unreviewed retrospective preprint, not a prospective clinical validation.
- Early-censored, lost-to-follow-up and mucinous cases were excluded from development, creating selection and scope limits.
- The ensemble was selected post hoc from 50 candidates and only matched the strongest external baseline within overlapping uncertainty.
- Morphology labels came from a frozen encoder and may contain label noise.
- No treatment decision, calibration study, workflow benefit or patient outcome was tested.
What to watch next
- Peer review and public release of the promised code
- Locked multi-hospital validation with calibration and subgroup error reporting
- Comparison with complete contemporary clinical and molecular risk models
- Prospective evidence that use changes decisions and improves patient outcomes
Living evidence record
Impact record IAI-0AMGER1
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
Whether morphology-aware learning from routine colorectal pathology slides can separate five-year survival risk across independent cohorts
Topics: Cancer · Digital pathology · Survival prediction · Clinical validation · Machine learning · Health equity
Regions: United States · United Kingdom · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Can routine tissue slides reveal a tumour’s molecular subtype?
A graph-constrained model separated two pancreatic-cancer subtypes reasonably well when transcriptomic signals were clear, but performance weakened sharply in 361 ambiguous cases and did not improve overall ranking over simpler image baselines.
9 min · 2 sources
Health & Life Sciences
Can AI diagnose ADHD reliably yet?
A 54-study meta-analysis found pooled sensitivity of 87% and specificity of 91%, but heterogeneity exceeded 96%, publication bias affected sensitivity, and recurring design weaknesses make clinical portability uncertain.
7 min · 1 source
Health & Life Sciences
Can AI make tumour mitosis counts more consistent?
A preprint study paired 13 pathologists' unaided and AI-assisted reviews of 385 tumour slides from three European centres. Agreement rose and counting time fell, but the unreviewed study did not establish which counts were correct, whether diagnoses improved or whether patients benefited.
11 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.