(1095-B) Step Up Your Morphological Cell Profiling Game with AI

Abstract: Background and Motivation
Reproducible morphological profiling, particularly for drug discovery, has become an important tool for compound evaluation. Here, we describe a novel approach for cell profiling using a fully automated computer vision approach with IKOSA AI. Our main motivation is to make Cell Painting easily accessible to biologists without any computer science background.
CellProfiler and CellPose are popular free, open-source software packages for image segmentation. CellProfiler is the current state-of-the-art for Cell Painting, it uses image processing pipelines that require manual parameter adjustments. CellPose is a generic deep learning-based segmentation model, but not particularly applicable to Cell Painting.
The primary objective of this work is to create a parameter-free, robust deep learning-based computer vision algorithm to achieve more precise object segmentations for fully-automated high-content imaging.

Materials and Methods
Our method consists of the following steps: we (1) generated ground truth labels from the Cell Painting Gallery JUMP-CP pilot dataset, (2) trained a Deep Learning segmentation model on a well-defined subset, (3) extracted features, and (4) compared them with CellProfiler output to validate our approach.
To create a representative dataset, we first determined an image subset to maximize the diversity of phenotypes. It includes different plates with pre-defined well positions, with eight channels per image, and was split into training (n=2,224 images), and validation (n=456 images). Ground truth labels for cell and nucleus were created by applying connected component analysis on CellProfiler-generated object outlines. All data was provided online in the IKOSA platform. Then, we trained a deep neural network using IKOSA AI and evaluated the segmentation performance on the validation dataset.
Segmentation outputs were further post-processed to reduce split-and-merge errors before extracting 3,664 features. To validate our feature extraction implementation, we compared them with CellProfiler output based on the original JUMP_analysis_v3 pipeline by correlation analysis. Feature compression was applied using StratoMineR for feature prioritization and to remove redundancy.

Results
We report promising results for image segmentation and feature extraction. Precision/recall/average precision for segmented cell and nuclei instances: 0.93/0.81/0.76, and 0.98/0.94/0.92; extracted features correlate with the CellProfiler output. The average segmentation runtime per image (1080x1080 pixels) is 2.2 seconds for our approach. A subset of 1,145 features was prioritized.

Discussion and Conclusions
This work presents a novel approach to Cell Painting image analysis that facilitates an easy-to-use, fully-automated feature extraction from raw image data without requiring programming, special hardware, or software. This method can be transferred to other morphological profiling datasets by model re-training. It will become available on the IKOSA platform for high-throughput scenarios and is expected to significantly impact resource efficiency in drug discovery and personalized medicine.