Automated detection of intracellular organelles with nnU-Net

In this page I will demonstrate the performance of nnU-Net, a deep learning-based segmentation method, in detecting intracellular organelles from electron microscopy (EM) images.

I chose to train the nnU-Net model by labelling mitochondria on EM images but the model is likely to perform very well on segmenting other intracellular organelles, as it has been shown to produce state-of-the-art segmentations on biomedical images of different kind.

Why nnU-Net

nnU-Net is a publicly available tool which is relatively easy to implement and has achieved a very high performance on many biomedical segmentation tasks. The fact that makes nnU-Net so interesting is its ability to adapt and find automatically the best configuration for different kinds of image data. This means that one does not need to be a machine learning expert to use it, and even machine learning experts can save a lot of time using this tool.

If you are interested to know more about nnU-Net look at: https://www.nature.com/articles/s41592-020-01008-z

Mitochondria segmentation with nnU-Net

Here an example on how well a nnU-Net trained model can produce segmentation of mitochondria. All mitochondria in the image, beside the two pointed by the arrows, are correctly predicted (green).

Note that this prediction was obtained without optimizing the training length and with a very little training dataset. The following two steps may therefore improve the model further.

Optimization of the training length

The model was trained for a few hours and it looked like its performance reached a plateau on a validation dataset, still the results shown here are a first trial and the training length was not optimized.

Larger training dataset

The model was trained on as few as 15 images of the same kind. It is very likely that using a larger training dataset may further reduce or eliminate the few errors.