**Background:** Predicting recurrence in low-grade, early-stage endometrial cancer (EC) is clinically challenging, as 5–10% of patients experience recurrence despite good overall prognosis. Traditional histology and molecular subtyping have limited predictive accuracy in this subgroup. Multiplexed immunofluorescence (IF) enables simultaneous visualization of multiple cellular markers, and when combined with deep learning, may uncover complex tumor-immune interactions that predict outcomes. This study aimed to develop a weakly-supervised deep learning framework (NaroNet) that could predict recurrence from multiplexed IF images without requiring manual expert annotation.
**Methods:** The study included 250 patients with low-grade (G1–G2), early-stage (FIGO I–II) endometrioid endometrial carcinoma from a retrospective collection at University Hospital La Paz (Madrid, Spain). Mean patient age was 64.5 years; 32 tumors recurred over a follow-up period of 30.9 months (IQR 18.3–50.5) after surgery. Most patients were FIGO stage IA and G1 without lymphovascular invasion; 62% received no adjuvant radiotherapy. Molecular subtypes included 45 (18%) mismatch repair protein-deficient and 8 (3.2%) POLE-mutated tumors; 96.8% showed wild-type p53 expression. A total of 489 TMA tumor cores (1.2 mm each; two cores per patient from different areas) were stained using a validated seven-color multiplexed IF protocol for simultaneous detection of CD68+ macrophages, CD8+ T cells, FOXP3+ regulatory T cells, PD-L1, PD-1, cytokeratin (CK, tumor cells), and DAPI. Staining was performed at two institutions (University of Navarra and Lunaphore Technologies) using the LabSat platform. Images were preprocessed with background subtraction (rolling ball algorithm, ImageJ). NaroNet, a weakly-supervised deep learning framework, was applied to learn tumor-immune interrelations at three levels: local phenotypes (20×20 µm patches), cellular neighborhoods (100×100 µm via graph neural networks with two hops), and tissue areas (~1800×1800 µm). Self-supervised learning (patch contrastive learning) embedded pixel information into 256-dimensional representation vectors. Hyperparameter tuning (400 architectures tested) identified optimal configuration: 10 local phenotypes, 8 cellular neighborhoods, 4 tissue areas, 2 hops, softmax activation, and constrained patch assignment. A tenfold cross-validation strategy was used (225 patients training, 25 patients testing per fold).
**Key Results:** The model achieved an AUC of 0.90 (95% CI: 0.83–0.95) and overall accuracy of 90.40% (95% CI: 86.72–94.08). Model predictions resulted in concordance for 96.8% of cases (κ=0.88). At the local phenotype level, three phenotypes with high immune cell infiltration (P1: FOXP3+CD68+; P3: CD8+CK+; P5: CD8+PD-1+) were significantly associated with no recurrence (all P<0.001). Phenotype P2 (tumor cells alone with variable PD-L1, no immune infiltration) was associated with recurrence (P=5.95e-28). At the cellular neighborhood level, N5 (tumor cells with no immune infiltration) was associated with recurrence (P=1.57e-29), while N3, N7, and N8 (showing immune infiltrates) were associated with no recurrence (all P<0.001). At the tissue area level, A1 and A4 ('cold tumors,' little immune infiltration) were associated with recurrence (P=2.22e-29 and P=2.04e-12, respectively), while A2 and A3 ('hot tumors,' high immune infiltration) were associated with no recurrence (P=5.93e-16 and P=6.25e-31, respectively). The NaroNet model (AUC=0.90) outperformed models based on mismatch repair protein status (AUC=0.79) and POLE mutation status (AUC=0.78). In robustness testing across two institutions, despite intensity variations (most significant for PD-L1, P<0.0001), overall quantification of relevant microenvironmental elements showed good correlations, and model predictions resulted in concordance of 92% and 85% in each institution separately.
**Clinical Implications:** This weakly-supervised deep learning model accurately predicts recurrence risk in low-grade, early-stage endometrial cancer using only two TMA cores from primary tumor resection, outperforming current molecular subtyping approaches. The model's interpretability—identifying specific local phenotypes, cellular neighborhoods, and tissue areas associated with recurrence—provides biologically meaningful insights. The finding that 'cold' (non-inflamed) tumors are strongly associated with recurrence while 'hot' (inflamed) tumors are associated with favorable outcomes suggests that immune microenvironment assessment could guide adjuvant therapy decisions. The decision tree analysis identified tissue area A1 ('cold tumors') as the most relevant feature for differentiating recurrence risk. Two misclassified patients who developed very late recurrences (86 and 90 months) highlight the complexity of tumor biology and suggest that additional factors beyond the seven-marker panel may be needed. The model's robustness across institutions supports potential clinical translation, though future work with larger independent cohorts and additional cellular markers is warranted.