A retrospective TMA-based study of digital pathology signatures for prognostic stratification in cervical cancer

Digital pathology may support prognostic stratification in cervical cancer, but evidence integrating routine histology with clinicopathologic variables remains limited. This retrospective study used a commercial cervical cancer tissue microarray comprising 126 cases. After quality-control exclusions, 121 cases were divided into a training cohort ( n = 85) and an internal test cohort ( n = 36). Digitized hematoxylin and eosin images were analyzed using a weakly supervised ResNet50-based pipeline. Sixty-four case-level pathology features were reduced by LASSO – Cox modeling to derive endpoint-specific signatures for overall survival (OS) and progression-free survival (PFS). A final nomogram combined age with the pathology signature. An independent 20-case cohort was used only for a directionally supportive assessment. The primary cohort included 28 OS events and 41 PFS events. Seven pathology features were retained for OS and six for PFS. In the internal test cohort, the nomogram achieved C-indices of 0.855 (95% CI, 0.742–0.968) for OS and 0.809 (95% CI, 0.693–0.925) for PFS. The nomogram performed better than the pathology-only model, whereas its improvement over the age-based clinical model was not statistically significant. Calibration and decision-curve analyses were exploratory because of the limited number of events. Estimates from the independent cohort were imprecise and were not used to support claims of transportability or clinical utility. The categorical immune-response scoring system was not independently prognostic. Pathology-derived image features may provide prognostic information in cervical cancer. However, the low event count, high feature burden, incomplete computational provenance, TMA sampling, and limited independent assessment preclude clinical application. Validation in larger multicenter whole-slide cohorts using a locked and reproducible pipeline is required.