Home

1. What is DenoiSeg?
2. How to use DenoiSeg?
3. Qualitative Results
4. Quantitative Performance Comparisons
5. Contact us

What is DenoiSeg?

DenoiSeg is a deep learning based segmentation method which can leverage unsupervised denoising with Noise2Void for cell/nuclei segmentation using only limited amount of ground truth segmentation annotations. The key contribution is to train a network for performing denoising and segmentation jointly. The reason for the success of DenoiSeg is that segmentation can profit from denoising, especially when performed jointly within the same network. The network becomes a denoising expert by seeing all available raw data, while co-learning to segment, even if only a few segmentation labels are available. We believe that DenoiSeg offers a viable way to circumvent the tremendous hunger for high quality training data and effectively enables few-shot learning of dense segmentations. For meow details, please see the paper.

How to use DenoiSeg?

DenoiSeg is available as a pip installable package. Just follow these simple steps here to get DenoiSeg running on your machine. We also provide example Jupyter notebooks to try DenoiSeg for some publicly available datasets from our publication.

Qualitative Results

We show results on three diverse datasets: DSB 2018, a developing Fly Wing and a Mouse Nuclei dataset. For each dataset, we add Gaussian noise with mean 0 and standard deviation 10 and 20. The dataset names are accordingly extended by n0, n10 and n20 to indicate the amount of additional noise. We compare against two baselines: (i) DenoiSeg trained purely for segmentation (referred to as Baseline), and a sequential scheme based on Prakash et. al that first trains a denoiser and then the aforementioned baseline (referred to as Sequential).

The row below depicts results for all noise levels for the case when only 10, 2 and 2 ground truth segmentation annotations for DSB, Fly Wing and Mouse Nuclei dataset respectively are available for training.

The row below depicts results for all noise levels for the case when only 38, 7 and 5 ground truth segmentation annotations for DSB, Fly Wing and Mouse Nuclei dataset respectively are available for training.

The row below depicts results for all noise levels for the case when only 304, 29 and 18 ground truth segmentation annotations for DSB, Fly Wing and Mouse Nuclei dataset respectively are available for training.

Quantitative Performance Comparisons

We compare the results of DenoiSeg with the above mentioned baselines in terms of AP score and SEG score. For each dataset and each noise level, the tables below present DenoiSeg setups with different values of hyperparameter $\alpha$. Additionally we also compare to results that use (the a priori unknown) best $\alpha$. The best $\alpha$ for each trained network is found by a grid search for $\alpha \in {0.1, 0.2, \dots, 0.9}$. The figures corresponding to the tables below can be found in the paper.