Cem M. Deniz, Ph.D.

Developing multimodal deep learning architectures and survival analysis models to forecast disease trajectories and surgical endpoints in knee osteoarthritis. Leveraging large-scale cohorts (OAI, MOST, and clinical data), our team pioneers vision transformer models (MR-Transformer), 3D volumetric CNNs, and foundation model adaptation (nnU-Net, MedSAM2) to predict total knee replacement (TKR) surgery within 3–10 year horizons and analyze time-to-pain progression.

Building self-supervised foundation models that integrate unstructured clinical text with medical imaging. Our work spans vision-language pretraining aligning narrative radiology reports with radiographs (MIDL, ML4H @ NeurIPS), as well as generative sequential pretraining (RAVEN) over electronic health records (EHRs) across 1M+ patient visits to achieve robust zero-shot disease onset forecasting while preserving diagnostic knowledge.

Overcoming the severe physical constraints of ultra-high field (7T+) magnetic resonance imaging through advanced parallel radiofrequency transmission (pTx) and computational electromagnetics. We design multi-channel RF transmit coil arrays, real-time B1+ shimming algorithms, 2D RF excitation pulse synthesis, and patient-specific Specific Absorption Rate (SAR) monitoring to maximize image homogeneity while maintaining strict clinical safety.

Bridging computational innovation into reproducible clinical radiology workflows through open-source software and international challenge benchmarks. We develop and distribute open-access MATLAB/Python GUIs for RF pulse design and RF shimming, while co-organizing international segmentation and prognosis challenges (KNOAP, IWOAI) to set rigorous standards for medical AI validation.