NUMERO: #20260186

Título: DEEP-HALO: Multimodal Deep Learning for Survival Prediction in Head and Neck Lymphomas via Spatial Connectomics

Nome do Apresentador: Lucas Lacerda de SOUZA

Categoria do Trabalho: Pôster - Investigação Original (IO)

Área Temática: Patologia Oral

Resumo: Objective: To overcome prognostic instability of classical statistical models in rare, heterogeneous malignancies by developing and validating an explainable multimodal deep learning framework (DEEP-HALO) integrating nuclear morphometry and spatial cellular connectomics for survival prediction in head and neck (H&N) lymphomas.
Study design: This retrospective international multicentre study analysed 578 confirmed cases of H&N lymphoma. A diagnostic-to-prognostic pipeline used CellViT for automated nuclear segmentation and feature extraction, and CellGNN to quantify spatial cellular interactions. Survival modelling was performed with DeepSurv, a non-linear neural network, and compared with penalised Cox regression across five random seeds to ensure robustness. Model explainability was assessed using SHAP to identify biological drivers.
Results: Classical Cox models showed instability and poor generalisation, producing unreliable hazard ratios and non-significant stratification. In contrast, DEEP-HALO remained stable, with consistent performance (mean C-index 0.58; mean IBS 0.34). In validation splits, the model reached a C-index of 0.79 and identified a high-risk group with significant separation (HR = 6.25; p = 0.049) and median survival of 4 months. SHAP analysis identified nuclear eccentricity, short inter-nuclear distance, and high Ki-67 expression as key features.
Conclusion: DEEP-HALO shows that deep learning can overcome heterogeneity limiting classical statistics, supporting spatial digital pathology for precision prognosis.

Autor 1:  Lucas Lacerda de SOUZA

E-mail 1:  [email protected]

Autor 2:   Hélder Antônio Rebelo PONTES

E-mail 2:  [email protected]

Autor 3 :  Alan Roger SANTOS-SILVA

E-mail 3:  [email protected]

Autor 4:  Marcio Ajudarte LOPES

E-mail 4:  [email protected]

Autor 5:  Felipe Paiva FONSECA

E-mail 5:  [email protected]

Autor 6:  Pablo Agustin VARGAS

E-mail 6:  [email protected]

Autor 7:  Syed Ali KHURRAM

E-mail 7:  [email protected]