Kidney Week 2025 Early Program - Advances in Research Conference 2025: Generative Artificial Intelligence in Kidney Disease Research and Management-From Pixels to Prognosis: The Role of Generative AI in Pathology

Video Transcription

Video Summary

Pinaki Sardar presents an educational overview of generative AI in kidney pathology, explaining how computers represent images as numbers (“embeddings”) in a latent space rather than human-interpretable descriptors (e.g., basement membrane thickness, luminal fraction). Traditional AI moves from images to features to patterns; generative models can go the reverse direction—starting from noise, navigating latent space, and producing realistic synthetic pathology images. He outlines key model types (AI → machine learning → deep learning → foundation models → large language/vision models) and describes self-supervised training that learns general image representations from large datasets.<br /><br />He highlights practical pathology use cases: stain transfer (e.g., converting H&E or autofluorescence images into PAS-like images), preserving tissue by generating virtual stains, predicting cell types/states from histology using paired molecular spatial data, and “label factories” that create synthetic images plus annotations to reduce expert labeling burden and improve segmentation performance.<br /><br />He emphasizes multimodal AI and vision-language models that combine slides with text (reports, clinical history) to answer questions and localize findings, citing large-scale cancer efforts (e.g., models trained on millions of whole-slide images). He also warns about limitations: rare/unseen events won’t be learned, synthetic images may lack biological validity, and current models can ignore images and hallucinate, making them unreliable for report-writing without careful validation. In discussion, he predicts AI will triage simpler cases, enabling pathologists to focus on complex ones, but stresses human–AI collaboration and multidisciplinary training.

Asset Subtitle

Pinaki Sarder

Meta Tag

Module ARC

Speaker Pinaki Sarder

Keywords

generative AI

kidney pathology

latent space embeddings

self-supervised learning

virtual staining

stain transfer

vision-language models

synthetic image annotation