Estimation of overall survival in immunotherapy-treated bladder cancer using computer vision.

Abstract

862 Background: Bladder cancer (BC) is a difficult and expensive cancer to treat, as lifelong care is needed. Anti-PD-1 therapies are used as a frontline treatment for BC. However, diverse response creates a need for novel methods to identify those patients that will benefit. We developed an algorithm (MIA-IO: Microscopy Image Analysis - Immunotherapy Outcome) to infer BC patient survival under anti-PD-1 immunotherapies with H&E-stained whole slide images (WSIs). Methods: A foundation model was trained on ~100k WSIs from various sources. Next, it was fine-tuned to create MIA-PDL-1, which infers PDL-1 status from H&E WSIs ( n = 1546). It was separately fine-tuned to identify tumor, stroma, necrosis and lymphocytes (training n = 2765 patches). An open-source model (CellViT) was used to identify neoplastic, inflammatory, connective, epithelial, and dead cells (training n = 190K cells). 617 features were derived from the segmentations to measure the spatial makeup of the tissues. Finally, a random survival forest (MIA-IO) was trained combining PDL-1 status, tissue/cell features, age and sex to estimate overall survival from treatment start. To develop MIA-IO, we used 331 WSIs from BC patient treated with several immunotherapies from internal trial and commercial sources (internal dataset) and 227 from a vendor as an external test set. We also evaluated MIA-IO using only patients treated with PD-1 therapies (i.e., Pembrolizumab and Nivolumab). Log-rank test was used to determine statistical significance of inferred high-risk (HR) and low-risk (LR) event times for each model. Results: The table shows survival probabilities for inferred HR and LR patient groups as stratified based on inferred death time. Survival of LR group was significantly higher than HR group for the all-immunotherapies model and for the PD-1 only therapy model in internal and external sets. Features holding greatest inference weight included texture of tumor regions and immune distributions. Cell and tissue features were found to carry the greatest estimation weight, such as texture of the tumor regions and immune distributions. Conclusions: Routine H&E biopsies may contain information prognostic of therapeutic response to immunotherapies and further study with more samples is warranted. This method is and faster than immunohistochemistry, spares additional tissue use, and may be used to select those patients that would benefit from PD-1 therapy. Survival probabilities based on algorithm stratification. Internal A Internal PD-1 External A External PD-1 Patients Train / Test ( n ) 264 / 67 157 / 39 - / 227 - / 182 MIA-IO LR / MIA-IO HR ( n ) 38 / 29 18 / 21 33 / 194 154 / 28 Overall Survival probability (LR / HR) 6-month 0.81 / 0.5 0.95 / 0.63 0.82 / 0.78 0.84 / 0.68 12-month 0.81 / 0.37 0.95 / 0.39 0.78 / 0.58 0.64 / 0.4 24-month 0.59 / 0.3 0.75 / 0.29 0.62 / 0.37 0.45 / 0.23 Log-rank test p -value 0.008 0.004 0.050 0.0182 A All patients, PD-1 PD-1 only treated cohort.