Teaching

Generative AI for Bioimage Analysis (GenAI4BIA)

Bridging Scales through Generative AI for Inverse Problems in Biomedical Computational Microscopy

Course Overview

This course covers the mathematical foundations of generative AI, inverse problems, and distribution learning in bioimage analysis, paired with a practical hands-on benchmark reproduction of VIRVS (Virus Infection Reporter Virtual Staining).

Topics Covered

Inverse Problems in Bioimaging: Forward modeling ($y = Ax + n$), ill-posedness, and Virtual Staining (mapping label-free brightfield micrographs to continuous fluorescence infection reporter signals).
Distribution Learning: Empirical distributions $p_{\text{data}}(x)$, Maximum Likelihood Estimation (MLE), KL divergence, Jensen-Shannon divergence, and Wasserstein distance.
Generative Model Families:
- Variational Autoencoders (VAEs): ELBO objective and reparameterization trick ($z = \mu + \sigma \odot \epsilon$).
- Generative Adversarial Networks (GANs / Pix2Pix): Minimax objective $\min_G \max_D V(D,G)$, conditional GAN loss, and L1 reconstruction.
- Diffusion Models (DDPM): Forward noise addition $q(x_t \mid x_{t-1})$, reverse process $p_\theta(x_{t-1} \mid x_t)$, and noise prediction MSE loss.
Code Implementation: Every mathematical concept is presented alongside its Python / PyTorch code equivalent.

Practical Workshop & Benchmarking

The hands-on component features interactive tutorials:

Data Download & Preparation: VIRVS dataset layout and RODARE access.
Autoencoders & VAEs: ELBO objective, generative sampling ($z \sim \mathcal{N}(0, I)$), and latent space interpolation.
U-Net Baseline Regression: Predicting continuous infection fluorescence from brightfield.
Pix2Pix Conditional GAN: Generative virtual staining model with adversarial loss.
Evaluation & Benchmarking: Metric evaluation (PSNR, SSIM, PCC, MAE) and cell-level viral reporter signal quantification.

For course slides (PDF), notebooks, and setup instructions, visit the GenAI4BIA GitHub repository.