AI-informed computational pathology classifier predicts outcomes across treatment modalities in muscle-invasive
Kamal Hammouda1, Naoto Tokuyama2, Germán Corredor3
1The Wallace H. Coulter Department of Biomedical Engineering at Georgia Tech and Emory University, Atlanta, GA, USA; Mathematics Department, Faculty of Science, Mansoura University, Mansoura, Egypt; Atlanta Veterans Affairs Medical Center, Atlanta, GA, USA.
Abstract:
Urothelial carcinoma (UC) is one of the leading causes of cancer-related mortality, and effective, scalable biomarkers for treatment planning remain limited. We present UC-TIL, an artificial intelligence (AI)-based model that quantifies spatial patterns of tumor-infiltrating lymphocytes (TILs) from routine H&E-stained slides to predict survival and immunotherapy response. We analyzed 558 whole-slide images across three cohorts: TCGA (D0&1, N = 292), Emory (D2, N = 161), and TRRC2819 (D3, N = 105), spanning chemotherapy and immune checkpoint inhibitor (ICI) treatments. UC-TIL classification was associated with OS (HR = 2.11, 95 %CI:1.01-4.41, p = 0.011) and PFS (HR = 3.68, 95 %CI:1.07-12.65, p = 0.0012) in locally advanced disease (D1 and D2), with consistent results in metastatic disease (D3) (HR = 1.73, 95 %CI:1.08-2.77, p = 0.043; PFS HR = 1.73, 95 %CI:1.07-2.81, p = 0.047). In the ICI-treated D3 cohort, UC-TIL achieved AUC = 0.757 and identified non-responders with 91 % specificity. UC-TIL enables reliable risk stratification and treatment response prediction in both locally advanced and metastatic urothelial carcinoma by analyzing spatial TIL patterns from standard pathology slides. These findings position UC-TIL as a readily deployable tool to guide personalized therapy across multiple clinical settings.
