Background
Several targeted therapies for FGFR alterations in bladder cancer are either currently in clinical
trials or already FDA-approved. FGFR alterations—including activating single nucleotide
variants (SNVs) and fusions—are common in bladder cancer and detectable via
next-generation sequencing of DNA and RNA. The ability to rapidly screen patients based on
routine pathology would help prioritize patients for full NGS workup. Here, we developed a
model using H&E whole slide images (WSIs) to predict FGFR alterations using real-world data.

Design
WSIs and ground truth labels pertaining to FGFR mutational status (obtained by DNA- and
RNA-seq) were collected from primary and metastatic bladder cancer specimens (N=3,706,
Table 1). Positive labels were defined as those harboring a pathogenic SNV or fusion of
FGFR, as confirmed by a molecular pathologist (n=214 FGFR3, n=10 other FGFR genes). A
holdout set (20% of data) was reserved for future validation purposes and not assessed
further. Model development was performed on the remaining set of data as follows: i) a
custom attention-based convolutional neural network with ResNet-18 backbone was trained
to predict FGFR alteration from each WSI in the training set (60%), ii) hyperparameters were
selected using an optimization set (20%) and iii) performance was reported on an evaluation
set of data (20%). Training, optimization, and evaluation was performed in 5-fold
cross-validation (CV). Cohorts were stratified to maintain a similar distribution of tissue site
and scanner types across each fold.

Results
In 5-fold CV, the FGFR model achieved a mean receiver operating characteristic area under curve (ROC-AUC) of 0.79 (95% CI 0.72 – 0.87, Figure 1). The ROC-AUC was 0.82 (95% CI
0.77 – 0.87) for WSIs within the bladder and 0.76 (95% CI 0.69 – 0.83) for non-bladder,
non-lymph node WSIs (Figure 2). The ROC-AUC was 0.80 (95% CI 0.72 – 0.88) for Philips
UFS WSIs and 0.78 (95% CI 0.68 – 0.87) for Leica GT450 WSIs. The ROC-AUC when
considering only FGFR3 mutated alterations was 0.79 (95% CI 0.72 – 0.87), and for FGFR1 /
FGFR2 alterations was 0.81 (95% CI 0.69 – 0.94).

Conclusion
This work demonstrates the ability to predict FGFR SNVs and fusions in bladder cancer using
a deep learning model trained on H&E WSIs. Performance was similar across tissue sites,
scanner types, and genes within the FGFR family. We anticipate that this model could be used
to rapidly identify patients enriched for FGFR alterations, who could then be prioritized for NGS
screening.

Table 1. Characteristics of the full cohort (N=3,706). A chi-square test was performed for each characteristic to determine association with the target label.

VIEW THE PUBLICATION

VIEW THE POSTER