Juan Carlos Caicedo

The paper develops a weakly supervised learning method to learn treatment effect representations from imaging data, improving downstream analysis performance.

The paper provides four datasets combining gene expression and morphology profiles from over 28,000 genetic and chemical perturbations for multi-modal analysis.

The paper compares morphology and gene expression data to determine their complementary roles in mapping cell states across drug perturbations.

This study uses deep learning and label-free imaging flow cytometry to classify red blood cell morphologies during storage, achieving 90% accuracy and predicting expiration dates based on morphological changes.

The paper introduces a deep learning framework for nucleus segmentation that uses image style transfer to adapt to new data without expert annotations, outperforming existing methods on diverse microscopy images.

This paper develops a weakly supervised CNN approach for learning single-cell feature embeddings from microscopy images, achieving better accuracy than classical features and competitiveness with transfer learning methods.

Adversarially learned features group fluorescence images of chemically treated human cells better than autoencoder features, though classical computer vision and transfer learning still do better.