Zero-Label Microscopy Anomaly Detection

The Problem

Pixel-reconstruction autoencoders fail on microscopy anomaly detection when corrupted images yield lower reconstruction error than sharp, high-contrast healthy cell membranes.

What It Does

• ✓Benchmarking 3 model paradigms (CAE-IF, SQAFormer, DINOv2 ViT embeddings) across 5,239 LIVECell frames
• ✓Empirical identification of catastrophic score sign inversion in reconstruction loss metrics
• ✓Feature extraction using self-supervised DINOv2 vision transformer backbones
• ✓Zero-shot metric evaluation including AUROC, AUPR, and Mahalanobis distance scoring
• ✓Anomaly visualization maps for cell image patches

How It Works & Architecture

Technical Challenges

• !Handling high-dimensional embeddings across 8 distinct cell lines without supervision.
• !Analyzing why mean squared error favors smooth blurred artifacts over sharp cell textures.

What I Learned

• ★Pixel MSE reconstruction error is an unreliable metric for biological visual anomaly detection.
• ★Self-supervised vision foundation models capture structure far better than task-specific trained autoencoders.