Learning a Latent Space of Cell Shapes via Signed Distance Functions
Cell shapes exhibit significant diversity, influenced by various biophysical forces and processes governed by the cell. Morphology is intricately linked to cellular function, with deviations from characteristic forms often indicating functional abnormalities. Analyzing cell shape is crucial for diagnosing various medical conditions, including cancer, red blood cell diseases, and neurological disorders. Recent advancements in microscopy and automation technologies enable large-scale experiments, generating millions of cell images daily. Consequently, morphological profiling has become an essential tool for extracting relevant shape information from microscopy images and facilitating new analyses based on similarities and differences among chemical or genetic perturbations. This study aims to examine cell shapes using natural images to obtain morphological profiles from fluorescence microscopy images of cultured cells. Our approach seeks to bridge the gap between classical image processing and transfer learning methods by introducing a novel method that encodes cell shape contours using a Signed Distance Function (SDF). This representation captures the shape boundary as the zero-contour of the learned function and implicitly represents the cell structure and biologically meaningful shape information. Our pipeline requires extracting contours through segmentation, similar to the classical machine learning approach. An Auto-Decoder architecture is employed to reconstruct cell contours via SDF and generate a low-dimensional representation of 2D shapes. By refining the hidden code words associated with each cell, the network learns to encode shapes from training data and retain shape information in a low-dimensional representation. Our proposed method offers three key benefits: it implicitly captures cell structure and biologically meaningful shape information, creates a mapping from the space of 2D shapes to a low-dimensional space for shape exploration, and generates a powerful shape descriptor for analysis applications. By leveraging these advantages, our approach offers a novel and effective method for encoding and working with 2D shape representations. This innovative method could provide more accurate and detailed shape information and have applications in cell analysis and drug discovery. Ultimately, our method aims to enhance our understanding of the underlying cellular behavior and the invasive potential of migrating cancer cells, potentially leading to new ways of controlling cell activity and advancing medical research.
S. Pieri, "Learning a Latent Space of Cell Shapes via Signed Distance Functions", M.S. Thesis, Computer Vision, MBZUAI, Abu Dhabi, UAE, 2023.