Table 4 Advantages and disadvantages of specific approaches for the performed image analysis tasks.
From: DeepBacs for multi-task bacterial image analysis using open-source deep learning approaches
Task | Network | Advantages | Disadvantages | Recommended for | Training speed |
|---|---|---|---|---|---|
(Instance) Segmentation | Classical U-Net | Better feature synthesis and correspondence with the input image when compared with classical fully connected neural networks. Reproducible inference in Fiji. | Requires annotated masks and postprocessing of the network output. | Low cell densities, high contrast, arbitrary cell shapes | Intermediate |
Multilabel U-Net | Semantic segmentation (background, cell boundary and cell cytosol) which improves to distinguish touching objects. Reproducible inference in Fiji. | Requires annotated masks and postprocessing of the network output. Implemented for 2D data. | Arbitrary cell shapes | Intermediate | |
StarDist | Highly generalisable and excellent performance at high object density; available for 2D and 3D; equipped for processing of large field of views; reproducible inference in Fiji, QuPath and Napari. | Limited to star-convex objects, does not work well for objects with large axial ratio (e.g., long rod-shaped cells). | Cocci, Ovococci, small rod-shaped bacterial cells (slow growth, stationary phase), all object densities | Fast | |
SplineDist | Regularly shaped, non-convex objects | Computationally expensive with a high demand of RAM memory; only implemented for 2D data. | Curved (non-star-convex) objects | Slow | |
Pix2pix | GAN-type architecture allows for arbitrary image-to-image translation tasks. | Longer training times, post-processing required; high demand of computational resources, risk of strong hallucinations; 2D. | Complex images with multimodal intensity distributions | Slow | |
Object detection | YOLOv2 | Fast training | Limited number of objects per image; low performance for small objects; fails determining objects in highly packed clusters; only available in 2D. | <50 uniformly distributed objects/image | Fast |
Denoising | CARE | Fast training for 2D and 3D data; the trained model can be deployed in Fiji. | Requires paired data (supervised network). | Targets that allow recording of low/high SNR data (slow or chemically fixed) | Fast |
Noise2Void | Unsupervised; new data is used both during the training and inference. Fast training; training and inference available in Fiji. | Lower performance than supervised learning approaches; only available for 2D. | Absence of high SNR images (fast dynamics, labels with low photostability) | (Very) fast | |
PureDenoise (parametric) | Multi-frame denoising; Fiji plugin; no special requirements and no training required. | Often lower performance than DL-based approaches. | Low SNR data with temporal correlation (e.g., processive movement) | N.A. | |
Artificial labelling | CARE | see above | Lower performance than fnet. | Prediction of membrane labels or structures visible in bright field images | Intermediate |
fnet | Training schedule and DL workflow is designed for artificial labelling | - | Prediction of membrane labels or structures visible in bright field images | Intermediate | |
Super-resolution prediction | CARE | see above | Might not predict rare sub-diffraction features | Regular structures (e.g., cell membranes) | Intermediate |
