Fig 1.
Schematic presentation of DeepMIB deep learning workflow.
(A) A scheme of a typical deep learning workflow. (B) Stepwise workflow of DeepMIB user interface for training or/and prediction. (C) Comparison of different padding types and overlapping strategies during prediction. Description for used images is given in Fig 1D. Arrowheads denote artefacts that may appear during prediction of datasets trained with the “same” padding parameter. (D) A snapshot of the GUI of the Activations explorer tool used for visualization of network features. Scale bar, 5 μm.
Fig 2.
Test datasets and acquired segmentation results.
(A) A middle slice of a serial section transmission electron microscopy dataset of the Drosophila melanogaster first instar larva ventral nerve cord supplemented with membrane prediction map and final segmentation of plasma membranes. (B) A random image of DNA channel from a high-throughput screen on human cultured osteocarcinoma U2OS cells (BBBC022 dataset, Broad BioImage Benchmark Collection) supplemented with prediction maps and final segmentation of nuclei, their boundaries (depicted in yellow), and interfaces between adjacent nuclei (depicted in red). The inset highlights a cluster of three nuclei. (C) A slice from a focus ion bean scanning electron microscopy dataset of the CA1 hippocampus region supplemented with prediction maps and 3D visualization of segmented mitochondria. In the right image, the blue box depicts area used for training and green box for testing and evaluation of the network performance. (D) A maximum intensity projection of 3D LM dataset supplemented with predictions and segmentation of inner hair cell located in the cochlea of the mouse inner ear. The inner hair cell cytoplasm depicted in vermillion, their nuclei in dark blue, and ribbon synapses in yellow. Nuclei of the surrounding cells are depicted in light blue. The light blue box indicates the area used for training, dark blue box for validation and green box (magnified in the inset) for testing and evaluation. The dataset was segmented using 3D U-net Anisotropic architecture, which was specially designed for anisotropic datasets. Scale bars, (A) 200 nm, (C) 1 μm, (D) 20 μm. (Scale bar for (B) not known). All presented examples are supplemented with movies (S1–S4 Movies) and DeepMIB projects including datasets and trained network (S1 Supplementary Material).