Inferring FGFR status from H&E images using digital pathology to identify patients for early-stage bladder cancer targeted therapies.

4593 Background: The identification of susceptible FGFR (Fibroblast Growth Factor Receptor) alterations may be critical in guiding treatment decisions for patients with bladder cancer. Current nucleic acid-based tests used to detect FGFR+ patients have limitations, including a slow turnaround time and high nucleic acid input requirement, especially in NMIBC, where tissue is often scarce. This study aims toevaluate the performance of an AI-based digital pathology algorithm, MIA:BLC-FGFR, adapted to investigate FGFR alterations in NMIBC patients from routine hematoxylin and eosin (H&E) stained whole slide images (WSIs). This approach may provide a rapid, low-cost, and effective alternative to nucleic acid testing. Methods: MIA:BLC-FGFR consists of an image quality control preprocessing stage, a Foundation Model (FM) pre-trained on ~55k unlabeled digital WSIs from various sources (multiple scanners, hospital systems, labs, diseases, tissue sites), and a classification module to enable inference of FGFR status from H&E-stained images. The classification module was trained on datasets (n = 3,067 WSIs) that included a mix of WSIs from multiple sources and disease stages (i.e., NMIBC, muscle-invasive and metastatic bladder cancer), and genetic classification provided by nucleic acid-based test. The algorithm was tuned to achieve a balanced specificity and sensitivity by selecting the operating point with highest F1 score (i.e., balanced sensitivity/specificity) in the training data. As part of this study, we then applied this model to WSIs of biopsies from 3 independent testing datasets (n = 578 WSIs) with varied NMIBC disease settings (i.e., high risk (HR) or intermediate risk (IR)) to evaluate the performance at predicting FGFR status, quantified by the Area Under ROC Curve (AUC). Results: MIA:BLC-FGFR demonstrated good concordance with nucleic acid testing methods. The results are summarized in the table below: Conclusions: The MIA:BLC-FGFR algorithm adapted to NMIBC can infer the presence or absence of select FGFR alterations from routine H&E images. This AI-based approach may offer a rapid, low-cost, and accurate alternative to traditional nucleic acid testing, particularly benefiting NMIBC patients with limited tumor tissue. By integrating into standard pathology workflows and providing results within minutes, the algorithm has the potential to significantly enhance FGFR testing rates and patient care decisions for emerging FGFR-targeted therapies. Testing datasets Independent Dataset 1 Independent Dataset 2 Independent Dataset 3 Disease setting HR NMIBC pT1 IR & HR NMIBC IR NMIBC Dataset size (FGFR+ %) 245 (29.7%) 163 (41%) 169 (49%) PPV 53% 64% 80% NPA 66% 71% 82% PPA 89% 73% 76% auROC 85% 80% 86%