Although recent therapeutic advances such as immunotherapy and the targeted agents enfortumab vedotin and erdafitinib have brought new treatment options, survival rates in advanced or metastatic disease have remained relatively stagnant. Efforts to better stratify patients for these targeted therapies have therefore taken center stage in both research and clinical practice.
One important molecular factor associated with urothelial carcinomas is alteration of the fibroblast growth factor receptor 3 (FGFR3). This receptor tyrosine kinase plays a vital role in controlling cellular growth and differentiation, and activating FGFR3 mutations or fusions are frequently observed in NMIBC, especially in papillary, low-grade lesions, where mutation rates can reach 80%. By contrast, advanced UC exhibits much lower rates, on the order of 10–15%. These data have fostered great interest in FGFR3 as a therapeutic target in urothelial cancer, culminating in the approval of erdafitinib for metastatic disease after other standard treatments have failed. However, given the relatively low prevalence of FGFR3 alterations in advanced-stage tumors, universal molecular testing can be inefficient and costly. Moreover, molecular testing approaches—hotspot PCR, RNA-based assays, or next-generation sequencing—often demand high-quality tissue, incur lengthy turnaround times, and can become cost-prohibitive if only one genetic locus is of interest. Consequently, there is a growing need for a practical means of pre-screening patients, enabling clinicians to identify those likely to harbor FGFR3-activating alterations, and thereby reduce unnecessary testing in those who do not.
In the study summarized here, a deep learning model was trained to detect FGFR3 mutations directly from digitized hematoxylin and eosin (H&E) sections of urothelial carcinoma. The investigators collected 1222 samples from patients who had NMIBC, MIBC, or metastatic urothelial carcinoma (mUC). Each tumor’s FGFR3 mutation status was defined by molecular assays such as SNaPshot PCR and next-generation sequencing, which pinpoint activating hotspot mutations. These data were then correlated with computational features extracted from the scanned H&E slides. Training took place on an initial discovery cohort consisting mainly of MIBC and NMIBC tumor images, and performance was validated on three independent cohorts comprising different clinical and pathological contexts: one from the publicly available TCGA (The Cancer Genome Atlas) bladder cancer dataset, another from a distinct set of MIBC patients (MIBC II), and a third from metastatic urothelial carcinoma patients (mUC). Despite the known molecular diversity between NMIBC and MIBC, including notable differences in morphological structure and mutational spectra, the final model achieved strong predictive results across all validation cohorts.
The deep learning pipeline employed a form of multiple-instance learning (MIL), in which each whole-slide image is subdivided into thousands of small patches (tiles). These tiles are first screened for tissue content and then fed into a self-supervised vision transformer network that extracts numeric descriptors reflecting morphological patterns. A secondary model aggregates tile-level features to yield a final prediction of whether FGFR3 is likely mutated or wild-type. Performance was measured by the area under the receiver operating characteristic curve (AUC), and the model proved both sensitive and accurate: AUC values were 0.82 in the TCGA MIBC cohort, 0.89 in an independent MIBC cohort, and 0.82 in the metastatic cohort. In practical terms, the high negative predictive value (approaching or reaching 99–100% in all validation sets) means that if the model suggests a tumor is wild-type, further molecular testing can reliably be skipped in about 40% of cases. That fraction of avoided tests could translate into significantly reduced expenditures and delays for patients and clinical laboratories, especially for those with advanced tumors in whom the yield of positive FGFR3 tests is relatively low.
Further analyses explored how well the model performed across different tumor subtypes, levels of inflammatory infiltration, and the presence of carcinoma in situ. The distribution of histological subtypes—conventional urothelial carcinoma, divergent differentiation (e.g., squamous or glandular), or neuroendocrine variants—had no major influence on results, with robust performance in all subgroups. The same consistency was observed whether tumor architecture was predominantly papillary or solid, and in tumors with or without carcinoma in situ. An exception was noted for low-grade versus high-grade carcinomas: the model showed an advantage in detecting FGFR3 mutations in low-grade tumors, which often display a more monomorphic, uniform urothelial cell pattern, suggesting that distinct architectural features or stromal reactions could be more readily recognized as signs of FGFR3 dysregulation.
An important insight emerged when examining outlier samples that were misclassified by the algorithm. For instance, tumors that carried activating FGFR3 mutations yet exhibited unusually low FGFR3 mRNA expression were more likely to resemble wild-type cases. Conversely, a subset of wild-type tumors with high FGFR3 expression looked deceptively like mutants. Excluding these atypical tumors improved the overall AUC substantially in the TCGA dataset, indicating that, from a morphological standpoint, FGFR3 activity—and not simply the presence of a mutation—drives the features that the network detects.
Pathologist review of the tiles that the model used to predict a mutant diagnosis highlighted a striking monomorphic tumor pattern, with minimal stroma or inflammatory infiltration, aligning with previous descriptions of FGFR3 mutant papillary lesions being relatively homogenous. Indeed, the least predictive tiles were characterized by higher pleomorphism, a greater degree of stromal desmoplasia, or pronounced inflammatory cell infiltrates, commonly found in non-FGFR3 mutant (wild-type) tumors. These observations also corroborated similar findings reported by other studies that visually linked the presence of FGFR3 alterations to more uniform nuclear morphology and low immune cell presence.
From a translational standpoint, these results underscore the promise of deep learning–assisted computational pathology in patient stratification. Since FGFR3 testing can incur additional costs and time (particularly for RNA-based assays that require high-quality templates), routine pathology slides could form the basis of a rapid triage method, flagging cases with a high probability of harboring FGFR3 alterations and sparing those with a negligible likelihood. This type of clinically focused model would ideally be deployed in real-time and integrated into the diagnostic workflow. With the increasing role of FGFR inhibitors—currently in the metastatic setting, but potentially also in higher-risk NMIBC—quicker and more efficient detection of FGFR3 abnormalities can translate into expedited treatment decisions and improved outcomes. The portability of the approach to other institutions or scanning platforms could be further enhanced by large-scale validation and calibration to standardize operating thresholds across different sample types and clinical conditions.
Although the results were encouraging, a few limitations merit emphasis. First, the sample size for rarer alterations such as FGFR3 fusions was too low to permit robust modeling. These patients are also candidates for erdafitinib, but gene fusions seem to be morphologically distinct—or simply too heterogeneous—and underrepresented to train the algorithm effectively. Second, the relatively small number of metastatic cases in which both primary and metastatic slides were available precluded more detailed comparisons of morphological shifts throughout disease progression. Future efforts could explore how well computational pathology models trained on one tumor stage generalize to tumors at other stages, as well as investigate strategies to incorporate immunohistochemistry or other molecular markers that capture the tumor microenvironment.
In conclusion, this study demonstrates that standard H&E slides contain reproducible morphological features predictive of FGFR3 mutation status in muscle-invasive and metastatic urothelial carcinoma. A deep learning model trained on a diverse set of UC samples achieved robust performance with high sensitivity and negative predictive value, making it well suited for a pre-screening or triage role in clinical diagnostics. By enabling rapid identification of likely FGFR3-mutant tumors, the model could curtail time spent and costs associated with molecular testing in patients who lack these alterations, focusing resources on those who stand to benefit from targeted FGFR inhibitors. Further confirmation in a large-scale, prospective manner would bolster clinical utility, and efforts to expand predictive capabilities to FGFR3 fusions or other novel targets would align with the broader trend of personalized therapeutics in bladder cancer. As computational pathology gains traction, these findings set the stage for a future in which automated image analysis complements standard molecular assays, yielding a more streamlined and accessible precision medicine approach for patients with urothelial carcinoma.
Written by: Markus Eckstein, MD, Institute of Pathology, University Hospital Erlangen & Bavarian Cancer Research Center, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany
References:
- Bannier, P.-A., Saillard, C., Mann, P., Touzot, M., Maussion, C., Matek, C., Klümper, N., Breyer, J., Wirtz, R., Sikic, D., Schmitz-Dräger, B., Wullich, B., Hartmann, A., Försch, S. & Eckstein, M. AI allows pre-screening of FGFR3 mutational status using routine histology slides of muscle-invasive bladder cancer. Nat. Commun. 15, 10914 (2024).
- Loeffler, C. M. L., Ortiz Bruechle, N., Jung, M., Seillier, L., Rose, M., Ghaffari Laleh, N., Knuechel, R., Brinker, T. J., Trautwein, C., Gaisa, N. T. & Kather, J. N. Artificial Intelligence-based detection of FGFR3 mutational status directly from routine histology in bladder cancer: a possible preselection for molecular testing? Eur. Urol. Focus 8, 472–479 (2022).
- Loriot, Y., Matsubara, N., Park, S. H., Huddart, R. A., Burgess, E. F., Houede, N., Banek, S., Guadalupi, V., Ku, J. H., Valderrama, B. P., Tran, B., Triantos, S., Kean, Y., Akapame, S., Deprince, K., Mukhopadhyay, S., Stone, N. L., Siefker-Radtke, A. O.; THOR Cohort 1 Investigators. Erdafitinib or Chemotherapy in Advanced or Metastatic Urothelial Carcinoma. N. Engl. J. Med. 389, 1961–1971 (2023).