MitoEM Dataset: Large-Scale 3D Mitochondria Instance Segmentation from EM Images

Automatic instance segmentation of mitochondria in electron microscopy data

“…Following acquisition and pre-processing of challenge data [1] (Sections 2.2, 2.3), we trained a series of customised 3D U-Nets (Sections 2.4, 2.6, 2.7) to predict mitochondria areas and boundaries (Section 2.8) in both the rat and human data. The predictions from these algorithms were combined in a multi-step post-processing procedure (Section 2.9), that resulted in high semantic and instance segmentation performance (Table 1).…”

“…Two datasets were provided for the development of 3D mitochondria instance segmentation approaches in association with the MitoEM Challenge [1]. The two 30 µm 3 image volumes provided were acquired from rat (Mito-R) and human cortex (Mito-H) (1000 × 4096 × 4096 in voxels at 30 × 8 × 8 nm resolution) with multi-beam scanning electron microscopy (for a full description of data acquisition see [1]).…”

“…For both volumes, 1000 consecutive slices were provided. Ground truth instance segmentation was generated for the two volumes through an iterative, semi-automated approach, that applied a 3D U-Net in collaboration with manual correction by expert annotators [1]. Ground truth mitochondria instance labels were provided for the first 500 slices, with the first 400 (0-399) used for training, and the remaining (400-499) used for validation.…”

“…The generation of both mitochondrial area and boundary predictions enabled application of a watershed-based method [1] to generate mitochondrial instance segmentations (Figure 1). To create seed areas for the marker-based watershed, the boundary mask was subtracted from the area mask to produce disconnected 3D 'islands'.…”

“…However, the development of novel analytical pipelines is often hampered by a scarcity of relevant, publicly available, well-labelled datasets. In response, the MitoEM dataset, two 3D mitochondria instance segmentations derived from rat and human cortex volume EM image stacks, were created and shared [1] along with a challenge to the communityto develop novel robust automated instance segmentation approaches for this data. Here, we present our response to this challenge.…”