⚖️ Transparency notice: this article was written with AI assistance and reviewed by the author, a medical oncologist.

The uncomfortable question before deciding

A patient with newly diagnosed colorectal cancer and no molecular result. The decision about immunotherapy may depend on one piece of data — the microsatellite instability (MSI) status — which today requires a test on tumor tissue: immunohistochemistry of MMR proteins or PCR. What if we could anticipate it with the images we already have, without waiting for the lab? That’s what a paper published in Academic Radiology attempts: a deep learning model that combines radiology and histology to non-invasively predict MSI.

The study

Researchers from two medical centers retrospectively analyzed 509 patients with pathology-confirmed colorectal cancer. The design follows the standard of medical AI studies:

• Center 1 (n = 379): divided into training (n = 261) and internal validation (n = 118)
• Center 2 (n = 130): independent external validation, with patients from another hospital
• Model: deep learning features extracted from multiphase CT (radiomics) and hematoxylin-eosin stained histological preparations (pathomics) using a pre-trained ResNet-101 network, plus clinical variables, integrated into a multiomic nomogram
• Interpretability: SHAP analysis to understand what each modality contributes to the prediction

Results: the multiomic nomogram makes the difference

Model	Training AUC	Internal validation AUC	External validation AUC
Multiomic nomogram (CT + histology + clinical)	0.996	0.999	0.993
Individual scores (radiomics, pathomics, preoperative)	0.919–0.963 (range across the three cohorts)
Clinical model	0.790	0.761	0.756

In the original study (PMID 42595634, DOI 10.1016/j.acra.2026.07.069), the nomogram integrating the three sources achieved an AUC of 0.996 in training, 0.999 in internal validation, and 0.993 in external validation, compared to 0.790/0.761/0.756 for the model using only clinical variables. In other words: adding imaging (CT + histology) to clinical data resulted in a huge leap in discriminative capacity — and the most important thing is that it held up in patients from a center that did not participate in training.

What does AI see that we don’t?

Here’s the conceptual key: the model is not reading mutations or sequences. It is detecting patterns — tumor texture on CT, gland architecture, cell density, stromal morphology in histology — that correlate with a distinct biological behavior. These are the same images that the radiologist and pathologist already report, but analyzed at a scale and with a consistency that the human eye cannot replicate. That the signal survives external validation suggests it is not just statistical noise.

Why it matters in clinical practice

MSI is not a laboratory whim: tumors with microsatellite instability (MSI-H) accumulate errors in repetitive DNA sequences due to a defect in repair systems (dMMR) and respond especially intensely to checkpoint inhibitor immunotherapy, regardless of the organ of origin. In colorectal cancer, MSI status guides from suspicion of hereditary syndromes (Lynch) to treatment choice in advanced stages. A non-invasive tool that anticipates that status with already available images could serve as preoperative triage — and help in scenarios where the sample is scarce or access to molecular biology is slow.

But with caution

• An AUC of 0.99 invites skepticism, not celebration. In a retrospective study, near-perfect figures may reflect overfitting, patient selection, or optimistic feature selection. External validation helps, but does not shield.
• Small cohort for deep learning: 509 patients total (261 for training) is little for a model of these characteristics. The signal is promising; generalization, still uncertain.
• Class imbalance: MSI represents a minority of colorectal cancers. A high AUC can coexist with modest sensitivity at the clinical cut-off — and that’s what matters in practice.
• Does not replace immunohistochemistry: IHC of MMR proteins and PCR are cheap, fast, and available in any pathology service. This AI is a triage hypothesis, not a substitute for the standard.
• Retrospective design and no real impact evaluation: we don’t know if it would change decisions or at what cost.

Reflection

This type of study represents the right direction: leveraging information we already pay for and already have (the CT, the slide) to advance a molecular decision. But the boundary between a promising result and a clinical tool is wide: prospective validation, direct comparison with the standard, and above all, demonstrating that it changes something for the patient are needed. Meanwhile, the practical message is clear — molecular biology still calls the shots, and AI is an ally under construction, not a substitute.

Reference

Ao W, Huang Y, Mao G, et al. Can Multiomics Modeling Enable Accurate Prediction of Microsatellite Instability in Colorectal Cancer? Academic Radiology (2026). PMID: 42595634 · DOI: 10.1016/j.acra.2026.07.069

— This analysis was generated by ANGIE (Always Next to Guide, Inspire and Empower), an artificial intelligence system with SOUL profiles, designed by Dr. Javier Pumares Pérez.

Connect on LinkedIn · ppj.es