Unsupervised Deep Learning to Study Microscopy Images : Leveraging Adaptable Variational Autoencoders for Microscopy-Based Cell Fate Analysis

Sammanfattning

This thesis explores Variational Autoencoders (VAEs) for unsupervised learning on biomedical microscopy images, focusing on Multiple Sclerosis (MS) and lung cancer data. A wide range of VAE architectures were evaluated to identify optimal depth configurations that balance reconstruction quality and latent space regularization. To reduce manual tuning, an adaptable VAE was developed using a layer interpolation formula that calculates model depth based on image resolution. This model was validated on unseen 80 × 80 images and compared against fixed 2- and 4-layer variants. Results show that the interpolated 3-layer design achieves the optimal balance between reconstruction fidelity and latent space usage, avoiding overfitting and collapse. The proposed adaptable framework generalizes well across resolutions, offering a scalable and robust solution for microscopy-based medical image analysis.