NUMERO: #20260550

Título: DIFFUSION MODELS AS AN EXPLAINABLE AI FRAMEWORK FOR ORAL POTENTIALLY MALIGNANT DISORDER PROGRESSION

Nome do Apresentador: Caíque Mariano PEDROSO

Categoria do Trabalho: Pôster - Investigação Original (IO)

Área Temática: Patologia Oral

Resumo: Objective: Diffusion models (DMs) are generative models that reconstruct images from latent embeddings, enabling visualization of feature-space transformations. Compared to generative adversarial networks, DMs provide more stable training and smoother latent-space interpolation. We evaluated their ability to generate biologically meaningful explanations of progression in oral potentially malignant disorders (OPMD).
Methods: A total of 2,157 OPMD whole-slide images were processed using Slideflow. Tiles were extracted at 224x124 pixels and feature embeddings were derived using the H0-mini foundation model (ViT-b/14, 86M parameters). CytoSyn, a latent DM based on the REPA-E architecture, decoded interpolated embedding trajectories into synthetic H&E images using GPU infrastructure (NVIDIA A100-SXM4-40GB, CUDA 12.5). We evaluated 5 latent trajectories with 10 interpolation steps (50 latent states). An expert oral pathologist assessed generate images according to the WHO OPMD grade.
Results: CytoSyn generated 3,200 synthetic H&E images across stochastic diffusion sampling. For visualization, trajectories were shown as 10-step interpolation strips. Expert evaluation consistently identified WHO-defined dysplastic features across stages, showing a gradual transition from non-progressor to progressor phenotypes with epithelial disorganization, basal polarity loss, nuclear hyperchromasia, mitoses, increased nuclear-to-cytoplasmic ratio, and pleomorphism.
Conclusion: DMs enable realistic and continuous visualization of disease progression, providing an interpretable framework for malignant transformation in OPMD.

Autor 1:  Caíque Mariano PEDROSO

E-mail 1:  [email protected]

Autor 2:   Siddhi RAMESH

E-mail 2:  [email protected]

Autor 3 :  Sara KOCHANNY

E-mail 3:  [email protected]

Autor 4:  Mark W LINGEN

E-mail 4:  [email protected]

Autor 5:  Pablo Agustin VARGAS

E-mail 5:  [email protected]

Autor 6:  Alan Roger SANTOS-SILVA

E-mail 6:  [email protected]

Autor 7:  Alexander T PEARSON

E-mail 7:  [email protected]