🇪🇸 Leer este artículo en español

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

Pathology is no longer just the pathologist’s domain

Muscle-invasive bladder cancer (MIBC) has a well-deserved reputation: two patients with the same stage and the same treatment can have completely different trajectories. After transurethral resection (TURBT) and before deciding between radical cystectomy with or without neoadjuvant chemotherapy, the oncologist faces an uncomfortable question: are we really stratifying these patients well with the information we have today? A US multicenter study, published in Urologic Oncology, proposes that the answer may be hidden in the H&E image we already have in the lab — and that an artificial intelligence platform can read it without the need for additional genomics.

What they did

The group led by Yair Lotan (UT Southwestern) and collaborators used an AI-powered computational histopathology platform (CHAI) that extracts histological features directly from whole-slide images of pre-treatment TURBT H&E preparations.

• Development: using The Cancer Genome Atlas (TCGA), they built a feature signature associated with the primary endpoint, recurrence-free survival (RFS). The continuous risk score was dichotomized into two groups: favorable and unfavorable.
• Validation: the model was “locked” and evaluated in an independent, retrospective, multicenter real-world cohort: 134 patients with cT2N0M0 urothelial carcinoma from NCI cancer centers who received radical cystectomy with or without neoadjuvant chemotherapy (NAC). Half (50%) received NAC.

Results: the H&E speaks of prognosis

In validation, the digital biomarker separated patients decisively (67 patients in each group):

• RFS: HR 3.1 (95% CI 1.7–5.7; p < 0.001) for the unfavorable group
• Cancer-specific survival (CSS): HR 3.5 (95% CI 1.5–7.8; p = 0.003)
• Overall survival (OS): HR 3.0 (95% CI 1.5–5.7; p = 0.001)

The most striking data point: 3-year RFS was 40% in the unfavorable group versus 74% in the favorable group (p < 0.001). And it’s not just that the signature works in crude form: after adjusting for clinical variables, including receipt of NAC, the biomarker remained associated with RFS, CSS, and OS (p < 0.01). Moreover, an exploratory analysis found a significant interaction between the biomarker and NAC for RFS (p = 0.02) — a hint, still a hypothesis, that this type of signature might one day help decide who truly benefits from neoadjuvant chemotherapy.

What does AI see that the human eye doesn’t?

Here’s the conceptual key: the platform is not reading mutations or gene expression. It is detecting morphological patterns — architecture, cell density, stromal features, subtle atypia — that correlate with more aggressive biological behavior. It’s the same H&E slide that the pathologist has already reported, but analyzed quantitatively, reproducibly, and at whole-slide scale.

💡 Why it matters (and what it means for MARTE)

For three reasons. First: it adds no cost or time — the input is the H&E that already exists, without an additional genomic panel. Second: it provides prognostic stratification beyond conventional pathology in a tumor where decisions (NAC yes or no?) remain difficult. Third, and more personal: it is exactly the type of evidence that supports MARTE’s philosophy — extracting clinical value from images already generated in daily practice, with models rigorously trained and validated. The methodology

— 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