You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
**A package for 3D cell segmentation with deep learning, including a napari plugin**: training, inference, and data review. In particular, this project was developed for analysis of mesoSPIM-acquired (cleared tissue + lightsheet) brain tissue datasets, but is not limited to this type of data. [Check out our preprint for more information!](https://www.biorxiv.org/content/10.1101/2024.05.17.594691v1)
12
14
13
-
- A napari plugin for 3D cell segmentation: training, inference, and data review. In particular, this project was developed for analysis of mesoSPIM-acquired (cleared tissue + lightsheet) datasets.
You can also generate docs by running ``make html`` in the docs/ folder.
27
+
You can also generate docs locally by running ``make html`` in the docs/ folder.
26
28
27
29
## Quick Start
28
30
@@ -37,6 +39,24 @@ Then go into Plugins > napari-cellseg3d, and choose which tool to use.
37
39
-**Train**: This module allows you to train segmentation algorithms from labeled volumes.
38
40
-**Utilities**: This module allows you to perform several actions like cropping your volumes and labels dynamically, by selecting a fixed size volume and moving it around the image; fragment images into smaller cubes for training; or converting labels from instance to segmentation and the opposite.
39
41
42
+
## Why use CellSeg3D?
43
+
44
+
The strength of our approach is we can match supervised model performance with purely self-supervised learning, meaning users don't need to spend (hundreds) of hours on annotation. Here is a quick look of our key results. TL;DR see panel **f**, which shows that with minmal input data we can outperform supervised models:
#### Performance of 3D Semantic and Instance Segmentation Models.
51
+
**a:** Raw mesoSPIM whole-brain sample, volumes and corresponding ground truth labels from somatosensory (S1) and visual (V1) cortical regions.
52
+
**b:** Evaluation of instance segmentation performance for several supervised models over three data subsets. F1-score is computed from the Intersection over Union (IoU) with ground truth labels, then averaged. Error bars represent 50% Confidence Intervals (CIs).
53
+
**c:** View of 3D instance labels from supervised models, as noted, for visual cortex volume in b evaluation.
54
+
**d:** Illustration of our WNet3D architecture showcasing the dual 3D U-Net structure with modifications (see Methods).
55
+
**e:** Example 3D instance labels from WNet3D; top row is S1, bottom is V1, with artifacts removed.
56
+
**f:** Semantic segmentation performance: comparison of model efficiency, indicating the volume of training data required to achieve a given performance level. Each supervised model was trained with an increasing percentage of training data (with 10, 20, 60 or 80%, left to right within each model grouping); Dice score was computed on unseen test data, over three data subsets for each training/evaluation split. Our self-supervised model (WNet3D) is also trained on a subset of the training set of images, but always without human labels. Far right: We also show performance of the pretrained WNet3D available in the plugin (far right), with and without removing artifacts in the image. See Methods for details. The central box represents the interquartile range (IQR) of values with the median as a horizontal line, the upper and lower limits the upper and lower quartiles. Whiskers extend to data points within 1.5 IQR of the quartiles.
57
+
**g:** Instance segmentation performance comparison of Swin-UNetR and WNet3D (pretrained, see Methods), evaluated on unseen data across 3 data subsets, compared with a Swin-UNetR model trained using labels from the WNet3D self-supervised model. Here, WNet3D was trained on separate data, producing semantic labels that were then used to train a supervised Swin-UNetR model, still on held-out data. This supervised model was evaluated as the other models, on 3 held-out images from our dataset, unseen during training. Error bars indicate 50% CIs.
58
+
59
+
40
60
## News
41
61
42
62
**New version: v0.2.0**
@@ -62,26 +82,10 @@ Previous additions:
62
82
63
83
64
84
65
-
### Install note for ARM64 (Silicon) Mac users
66
-
67
-
To avoid issues when installing on the ARM64 architecture, please follow these steps.
68
-
69
-
1) Create a new conda env using the provided conda/napari_CellSeg3D_ARM64.yml file :
3) Launch napari, the plugin should be available in the plugins menu.
79
-
80
-
81
85
82
86
## Requirements
83
87
84
-
**Python 3.8 or 3.9 required.**
88
+
**Compatible with Python 3.8 to 3.10.**
85
89
Requires **[napari]**, **[PyTorch]** and **[MONAI]**.
86
90
Compatible with Windows, MacOS and Linux.
87
91
Installation should not take more than 30 minutes, depending on your internet connection.
@@ -92,17 +96,22 @@ A CUDA-capable GPU is not needed but very strongly recommended, especially for t
92
96
93
97
If you get errors from MONAI regarding missing readers, please see [MONAI's optional dependencies] page for instructions on getting the readers required by your images.
94
98
95
-
## Quick demo
99
+
### Install note for ARM64 (Silicon) Mac users
100
+
101
+
To avoid issues when installing on the ARM64 architecture, please follow these steps.
96
102
97
-
After installation, you can run the plugin by running:
103
+
1) Create a new conda env using the provided conda/napari_CellSeg3D_ARM64.yml file :
0 commit comments