3D Mitochondria Shape Library for Optical Microscopy (3DMSL): A multimodal dataset for deep learning based mitochondrial analysis

TL;DR

3DMSL is a new dataset of mitochondria shapes designed to improve deep learning models for analyzing microscope images.

Contribution

3DMSL introduces a physics-based simulator to generate annotated 3D fluorescence microscopy datasets for training deep learning models.

Findings

3DMSL contains over 27,000 instances of mitochondria in various 3D shape formats.

The dataset supports applications like segmentation and 3D reconstruction from microscope images.

3DMSL enables microscope-to-microscope translation and time-lapse video creation with 3D ground truth.

Abstract

With the increasing development of deep learning solutions for fluorescence microscopy image analysis, there is a growing demand for annotated ground truth datasets to train supervised methods. However, obtaining these annotations is a laborious and expensive endeavor. To address this problem for microscope analysis of cell organelles, we release 3DMSL, a database of 3D shapes of mitochondria. 3DMSL utilizes high-resolution Electron Microscopy data as the source for creating the extensive database. Utilizing a physics-based simulator, 3DMSL enables the creation of large fluorescence microscope image datasets with 3D ground truths, which can be used to train deep learning models for various applications, including segmentation, 3D reconstruction from images and stacks, creation of time-lapse videos with 3D ground truth, and microscope-to-microscope translation. 3DMSL contains >27k instances of diverse mitochondria shapes in different 3D shape representation formats of meshes, point clouds and implicit shapes.

Figures