Abstract
Electron microscopy (EM) allows the identification of intracellular organelles such as mitochondria, providing insights for clinical and scientific studies. In recent years, a number of novel deep learning architectures have been published reporting superior performance, or even human-level accuracy, compared to previous approaches on public mitochondria segmentation datasets. Unfortunately, many of these publications make neither the code nor the full training details public, leading to reproducibility issues and dubious model comparisons. Thus, following a recent code of best practices in the field, we present an extensive study of the state-of-the-art architectures and compare them to different variations of U-Net-like models for this task. To unveil the impact of architectural novelties, a common set of pre- and post-processing operations has been implemented and tested with each approach. Moreover, an exhaustive sweep of hyperparameters has been performed, running each configuration multiple times to measure their stability. Using this methodology, we found very stable architectures and training configurations that consistently obtain state-of-the-art results in the well-known EPFL Hippocampus mitochondria segmentation dataset and outperform all previous works on two other available datasets: Lucchi++ and Kasthuri++. The code and its documentation are publicly available at https://github.com/danifranco/EM_Image_Segmentation .