An interactive deep learning-based approach reveals mitochondrial cristae topologies

Abstract

Outer and inner mitochondrial membranes are highly specialized structures with distinct functional properties. Reconstructing complex 3D ultrastructural features of mitochondrial membranes at the nanoscale requires analysis of large volumes of serial scanning electron tomography data. While deep-learning-based methods improved in sophistication recently, time-consuming human intervention processes remain major roadblocks for efficient and accurate analysis of organelle ultrastructure. In order to overcome this limitation, we developed a deep-learning image analysis platform called Python-based Human-In-the-LOop Workflows (PHILOW). Our implementation of an iterative segmentation algorithm and Three-Axis-Prediction method not only improved segmentation speed, but also provided unprecedented ultrastructural detail of whole mitochondria and cristae. Using PHILOW, we found that 42% of cristae surface exhibits tubular structures that are not recognizable in light microscopy and 2D electron microscopy. Furthermore, we unraveled a fundamental new regulatory function for the dynamin-related GTPase Optic Atrophy 1 (OPA1) in controlling the balance between lamellar versus tubular cristae subdomains.

Protocols for Generating Surfaces and Measuring 3D Organelle Morphology Using Amira

“…Amira is highly customizable through an animation creation tool, the ability to assign colors to various organelles, interpolation, compatibility with a wide range of import and export files, and simple workflows with either manual or semi-automated segmentation [ 32 , 33 ] ( Supplemental Figure S1 ). Further modification is possible through scripting interfaces using MATLAB and Python [ 29 ], which is important because deep-learning algorithms on Python can expedite the segmentation workflow [ 34 ]. The flexibility of Amira provides advantages for both beginners and experienced researchers over less costly open-source software, such as ImageJ.…”

Serial Block Face‐Scanning Electron Microscopy as a Burgeoning Technology

“…One prominent example is Python-based Human-In-the-LOop Workflows, developed by Suga and colleagues, which allow for deep-learning of mitochondrial and cristae networks for their rapid analysis. [39] Fully automated segmentation can significantly reduce analysis time and minimize human error, but the accuracy of the segmentation depends on the quality of the training data and the performance of the algorithm. For example, ideally, in a deep learning method, a user can simply input the data set into the algorithm and adjust parameters and verify the output of completed 3D reconstructions of selected organelles.…”

“…[38] Once validated, studies have found that the proofreading time required to verify machine-learning approaches is quite low. [11,39,40] Yet, many studies using fully-automated methods are employing these approaches for a limited set of organelles not yet expanding to the full spectrum of biopsies and pathophysiology. However, the diverse phenotypes of mitochondria, or other organelles, limit the ability of SBF-SEM machine learning.…”

Three-dimensional ultrastructure analysis of organelles in injured motor neuron

“…The main strategy for auto-segmentation is a three-dimensional convolutional neural network (CNN) architecture based on a three-dimensional U-Net (Çiçek et al 2016;Falk et al 2019). Some developing tools have been released by an open source, such as a repository providing large quantities of reliable data, codes, and trained models (Heinrich et al 2021), a new pipeline to train a CNN effectively (Gallusser et al 2023), improved DL platforms (Suga et al 2021), and so on. However, more challenges remain in an ongoing effort to reduce human labour for the generation of training data, proofreading predictions, and reducing computation costs.…”