Mining Typical Phenotypes from Cell Images Using Stable Diffusion

Abstract

This thesis studies whether diffusion-model typicality can improve cell selection for image-based morphological profiling. The project focuses on Broad Bioimage Bench mark Collection 021 (BBBC021), a Cell Painting-style benchmark with mechanism of-action (MoA) labels and strong batch effects. The evaluation uses a controlled selector benchmark where every method starts from the same border-filtered Cellpose SAM proposals, and only the rule for retaining proposals changes. The baseline keeps all valid proposals, random selection subsamples them, foreground fraction ranks crops by target-cell pixel coverage, and MorphoDiff-Typicality ranks the same proposals by a population-null MorphoDiff score computed over the target-cell mask. A fixed matched subset is used so that all headline comparisons operate on the same perturbations, wells, and images. The final selector benchmark uses isolated target-cell crops, removes border-touching proposals, embeds crops with InceptionV3, normalizes the resulting cell features with cell-level typical variation normalization (TVN), and evaluates well-level profiles with copairs. The baseline retains 389,448 crops and reaches perturbation mean average precision (mAP) 0.587, perturbation fraction retrieved (FR) 57.65%, MoA mAP 0.587, and MoA FR 91.67%. Foreground-fraction ranking gives the strongest MoA mAPresult, peaking at 0.635 with 20% retained crops. MorphoDiff-Typicality gives the strongest perturbation FR result, peaking at 62.35% with 45% retained crops, and reaches the MoA FR baseline at 35%. These results show that simple crop quality is a strong signal in BBBC021, while diffusion-based typicality preserves a different perturbation-aware signal that is most visible in perturbation-level detectability.