Abstract 3997: Elucidating broad cell-cell interactions via 3D super-cell graphs for prostate cancer risk stratification

Key Points

• The study aims to enhance prostate cancer prognostication through 3D super-cell graphs that capture complex cell interactions.
• Developed a slide-free 3D pathology method generating volumetric datasets mimicking H&E staining.
• Applied deep-learning segmentation (Cellpose) to extract cell nuclei and their histomorphometric features.
• Used Leiden clustering to group similar cells into supercells, preserving interactions with reduced complexity.
• Constructed 3D supercell graphs using k-nearest neighbors for a cohort of 76 prostatectomy specimens.
• Evaluated prognostic features using LASSO modeling with cross-validation to predict biochemical recurrence.
• Achieved an AUC of 0.745 ± 0.06 for predicting 5-year biochemical recurrence using 3D supercell graph features.
• 3D histomorphometric features from cancer glands yielded a similar AUC of 0.751 ± 0.08.
• Combining features resulted in a higher AUC of 0.837 ± 0.05, indicating strong predictive capacity.
• Explored additional prognostic value of end-to-end graph neural network learning in low-shot settings.