A1869
Title: Persistence image-based deep Cox modeling for MCI-to-AD conversion time under right-censoring
Authors: Somin Lee - Chungnam National University (Korea, South) [presenting]
Sunghun Kim - Korea university (Korea, South)
Eunjee Lee - Chungnam National University (Korea, South)
Bo-yong Park - Korea University (Korea, South)
Abstract: Early identification of Alzheimers disease progression risk is important for adjusting follow-up frequency, planning interventions and medications, and allocating care resources. In clinical cohorts, conversions are often not observed within the follow-up window; right-censoring is common, and standard binary classification (0/1) may be inadequate. A pipeline that derives topological features from hippocampal surface measurements and performs censoring-aware risk modeling via Deep Cox objectives is presented. Compared with using surface maps directly, topological summaries aim to capture multiscale structural patterns and provide fixed-dimensional representations for learning. Hippocampal radial distance is arranged into an image-like scalar field, and cubical-complex persistent homology is computed to obtain birth-death intervals (Persistence Diagrams, PD). PDs are vectorized into Persistence Images (PI) via Gaussian smoothing and grid integration, yielding fixed-dimensional inputs. Experiments were conducted on baseline MCI subjects (N=322; events=188; censored=134) using 5-fold cross-validation. MLP-DeepCox on flattened PI vectors achieved mean C-index 0.5714+-0.0131; patch-token ViT-DeepCox achieved 0.5545+-0.0335; RawCNN trained on raw surface maps without PI yielded 0.5400+-0.0430. Results are preliminary due to limited sample size and fold-to-fold variation, but suggest the PD-PI-Deep Cox pipeline may be useful for time-to-conversion risk ranking under right-censoring.
