⚖️ 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 results. The decision regarding 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 is what a paper published in Academic Radiology attempts: a deep learning model that combines radiology and histology to predict MSI non-invasively.

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 from hematoxylin-eosin stained histological preparations (pathomics) using a pretrained 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 a training AUC of 0.996, internal validation AUC of 0.999, and external validation AUC of 0.993, 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 point 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 is 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 on hist

— 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