The Final Information to nnU-Internet. A theoretical and sensible information on… | by François Porcher | Aug, 2023

Every thing you should know to grasp the State of the Artwork nnU-Internet, and how you can apply it to your personal dataset.

Neuroimaging, by Milak Fakurian on Unsplash, link

Throughout my Analysis internship in Deep Studying and Neurosciences at Cambridge College, I used the nnU-Internet quite a bit, which is an especially robust baseline in Semantic Picture Segmentation.

Nevertheless, I struggled slightly to completely perceive the mannequin and how you can practice it, and didn’t discover a lot assistance on web. Now that I’m snug with it, I created this tutorial that can assist you, both in your quest to grasp higher what’s behind this mannequin, or how you can use it in your personal dataset.

All through this information, you’ll:

  1. Develop a concise overview of the important thing contributions of nnU-Internet.
  2. Discover ways to apply nnU-Internet to your personal dataset.

All code obtainable on this Google Collab notebook

This work took me a major quantity of effort and time. If you happen to discover this content material helpful, please contemplate following me to extend its visibility and assist help the creation of extra such tutorials!

Acknowledged as a state-of-the-art mannequin in Picture Segmentation, the nnU-Internet is an indomitable power in terms of each 2D and 3D picture processing. Its efficiency is so sturdy that it serves as a robust baseline in opposition to which new laptop imaginative and prescient architectures are benchmarked. In essence, in case you are venturing into the world of growing novel laptop imaginative and prescient fashions, contemplate the nnU-Internet as your ‘goal to surpass’.

This highly effective software is predicated on the U-Internet mannequin (You’ll find one in all my tutorials right here: Cook your first U-Net), which made its debut in 2015. The appellation “nnU-Internet” stands for “No New U-Internet”, a nod to the truth that its design doesn’t introduce revolutionary architectural alterations. As an alternative, it takes the present U-Internet construction and squeezes out its full potential utilizing a set of ingenious optimization methods.

Opposite to many fashionable neural networks, the nnU-Internet doesn’t depend on residual connections, dense connections, or consideration mechanisms. Its energy lies in its meticulous optimization technique, which incorporates methods like resampling, normalization, considered alternative of loss perform, optimiser settings, knowledge augmentation, patch-based inference, and ensembling throughout fashions. This holistic method permits the nnU-Internet to push the boundaries of what’s achievable with the unique U-Internet structure.

Whereas it’d appear to be a singular entity, the nnU-Internet is in actual fact an umbrella time period for 3 distinct varieties of U-Nets:

2D, 3D, and cascade, Picture from nnU-Net article
  1. 2D U-Internet: Arguably essentially the most well-known variant, this operates immediately on 2D photographs.
  2. 3D U-Internet: That is an extension of the 2D U-Internet and is able to dealing with 3D photographs immediately by means of the appliance of 3D convolutions.
  3. U-Internet Cascade: This mannequin generates low-resolution segmentations and subsequently refines them.

Every of those architectures brings its distinctive strengths to the desk and, inevitably, has sure limitations.

As an example, using a 2D U-Internet for 3D picture segmentation might sound counterintuitive, however in follow, it might nonetheless be extremely efficient. That is achieved by slicing the 3D quantity into 2D planes.

Whereas a 3D U-Internet could appear extra refined, given its greater parameter rely, it isn’t at all times essentially the most environment friendly answer. Notably, 3D U-Nets typically battle with anisotropy, which happens when spatial resolutions differ alongside completely different axes (for instance, 1mm alongside the x-axis and 1.2 mm alongside the z-axis).

The U-Internet Cascade variant turns into significantly helpful when coping with giant picture sizes. It employs a preliminary mannequin to condense the picture, adopted by an ordinary 3D U-Internet that outputs low-resolution segmentations. The generated predictions are then upscaled, leading to a refined, complete output.

Picture from nnU-Net article

Sometimes, the methodology entails coaching all three mannequin variants inside the nnU-Internet framework. The next step could also be to both select the most effective performer among the many three or make use of ensembling methods. One such approach may contain integrating the predictions of each the 2D and 3D U-Nets.

Nevertheless, it’s price noting that this process may be fairly time-consuming (and in addition cash since you want GPU credit). In case your constraints solely enable for the coaching of a single mannequin, fret not. You’ll be able to select to solely practice one mannequin, because the ensembling mannequin solely brings very marginal positive factors.

This desk illustrates the best-performing mannequin variant in relation to particular datasets:

Picture from nnU-Net article

Dynamic adaptation of community topologies

Given the numerous discrepancies in picture dimension (contemplate the median form of 482 × 512 × 512 for liver photographs versus 36 × 50 × 35 for hippocampus photographs), the nnU-Internet intelligently adapts the enter patch dimension and the variety of pooling operations per axis. This primarily implies an computerized adjustment of the variety of convolutional layers per dataset, facilitating the efficient aggregation of spatial info. Along with adapting to the various picture geometries, this mannequin takes under consideration technical constraints, resembling obtainable reminiscence.

It’s essential to notice that the mannequin doesn’t carry out segmentation immediately on your entire picture however as a substitute on fastidiously extracted patches with overlapping areas. The predictions on these patches are subsequently averaged, resulting in the ultimate segmentation output.

However having a big patch means extra reminiscence utilization, and the batch dimension additionally consumes reminiscence. The tradeoff taken is to at all times prioritize the patch dimension (the mannequin’s capability) relatively than the batch dimension (solely helpful for optimization).

Right here is the Heuristic algorithm used to compute the optimum patch dimension and batch dimension:

Heuristic Rule for Batch and Patch Dimension, Picture from nnU-Net article

And that is what it seems like for various Datasets and enter dimensions:

Structure in perform of the enter picture decision, Picture from nnU-Net article

Nice! Now Let’s rapidly go over all of the methods utilized in nnU-Internet:


All fashions are skilled from scratch and evaluated utilizing five-fold cross-validation on the coaching set, that means that the unique coaching dataset is randomly divided into 5 equal elements, or ‘folds’. On this cross-validation course of, 4 of those folds are used for the coaching of the mannequin, and the remaining one fold is used for the analysis or testing. This course of is then repeated 5 occasions, with every of the 5 folds getting used precisely as soon as because the analysis set.

For the loss, we use a mixture of Cube and Cross Entropy Loss. It is a very frequent loss in Picture Segmentation. Extra particulars on the Cube Loss in V-Net, the U-Net big’s brother

Knowledge Augmentation methods

The nnU-Internet have a really robust Knowledge Augmentation pipeline. The authors use random rotations, random scaling, random elastic deformation, gamma correction and mirroring.

NB: You’ll be able to add your personal transformations by modifying the supply code

Elastic deformation, from this article
Picture from OpenCV library

Patch primarily based Inference

In order we stated, the mannequin doesn’t predict immediately on the total decision picture, it does that on extracted patches after which aggregates the prediction.

That is what it seems like:

Patch Based mostly inference, Picture by Creator

NB: The patches within the heart of the image are given extra weight than those on the facet, as a result of they include extra info and the mannequin performs higher on them

Pairwise Mannequin Ensembling

Mannequin Ensembling, Picture by writer

So when you bear in mind effectively, we are able to practice as much as 3 completely different fashions, 2D, 3D, and cascade. However after we make inference we are able to solely use one mannequin at a time proper?

Properly seems that no, completely different fashions have completely different strengths and weaknesses. So we are able to truly mix the predictions of a number of fashions in order that if one mannequin may be very assured, we prioritize its prediction.

nnU-Internet checks each mixture of two fashions among the many 3 obtainable fashions and picks up the most effective one.

In Apply, there are 2 methods to try this:

Laborious voting: For every pixel, we have a look at all the chances outputted by the two fashions, and we take the category with the best likelihood.

Tender Voting: For every pixel, we common the likelihood of the fashions, after which we take the category with the utmost likelihood.

Earlier than we start, you may obtain the dataset here and observe the Google Collab notebook.

If you happen to didn’t perceive something concerning the first half, no worries, that is the sensible half, you simply must observe me, and you’re nonetheless going to get the most effective outcomes.

You want a GPU to coach the mannequin in any other case it doesn’t work. You’ll be able to both do it domestically, or on Google Collab, don’t neglect to alter the runtime > GPU

So, to begin with, you should have a dataset prepared with enter photographs and their corresponding segmentation. You’ll be able to observe my tutorial by downloading this prepared dataset for 3D Mind segmentation, after which you may change it with your personal dataset.

Downloading knowledge

To start with you must obtain your knowledge and place them within the knowledge folder, by naming the 2 folders “enter” and “ground_truth” which incorporates the segmentation.

For the remainder of the tutorial I’ll use the MindBoggle dataset for picture segmentation. You’ll be able to obtain it on this Google Drive:

We’re given 3D MRI scans of the Mind and we wish to section the White and Grey matter:

Picture by Creator

It ought to seem like this:

Tree, Picture by Creator

Organising the principle listing

If you happen to run this on Google Colab, set collab = True, in any other case collab = False

collab = True

import os
import shutil
from collections import OrderedDict
import json
import numpy as np

#visualization of the dataset
import matplotlib.pyplot as plt
import nibabel as nib

if collab:
from google.colab import drive
drive.mount('/content material/drive', force_remount=True)
# Change "neurosciences-segmentation" to the title of your challenge folder
root_dir = "/content material/drive/MyDrive/neurosciences-segmentation"

# get the dir of the mother or father dir
root_dir = os.getcwd()

input_dir = part of(root_dir, 'knowledge/enter')
segmentation_dir = part of(root_dir, 'knowledge/ground_truth')

my_nnunet_dir = part of(root_dir,'my_nnunet')

Now we’re going to outline a perform that creates folders for us:

def make_if_dont_exist(folder_path,overwrite=False):
creates a folder if it doesn't exists
folder_path : relative path of the folder which must be created
over_write :(default: False) if True overwrite the present folder
if os.path.exists(folder_path):

if not overwrite:
print(f'{folder_path} exists.')
print(f"{folder_path} overwritten")

print(f"{folder_path} created!")

And we use this perform to create our “my_nnunet” folder the place the whole lot goes to be saved

make_if_dont_exist('my_nnunet', overwrite=False)
print(f"Present working listing: {os.getcwd()}")

Library set up

Now we’re going to set up all the necessities. First let’s set up the nnunet library. In case you are in a pocket book run this in a cell:

!pip set up nnunet

In any other case you may set up nnunet immediately from the terminal with

pip set up nnunet

Now we’re going to clone the nnUnet git repository and NVIDIA apex. This incorporates the coaching scripts in addition to a GPU accelerator.

!git clone
!git clone

# repository dir is the trail of the github folder
respository_dir = part of(my_nnunet_dir,'nnUNet')
!pip set up -e
!pip set up --upgrade git+

Creation of the folders

nnUnet requires a really particular construction for the folders.

task_name = 'Task001' #change right here for various activity title

# We outline all the mandatory paths
nnunet_dir = "nnUNet/nnunet/nnUNet_raw_data_base/nnUNet_raw_data"
task_folder_name = part of(nnunet_dir,task_name)
train_image_dir = part of(task_folder_name,'imagesTr') # path to coaching photographs
train_label_dir = part of(task_folder_name,'labelsTr') # path to coaching labels
test_dir = part of(task_folder_name,'imagesTs') # path to check photographs
main_dir = part of(my_nnunet_dir,'nnUNet/nnunet') # path to foremost listing
trained_model_dir = part of(main_dir, 'nnUNet_trained_models') # path to skilled fashions

Initially the nnU-Internet was designed for a decathlon problem with completely different duties. When you have completely different duties simply run this cell for all of your duties.

# Creation of all of the folders
overwrite = False # Set this to True if you wish to overwrite the folders
make_if_dont_exist(task_folder_name,overwrite = overwrite)
make_if_dont_exist(train_image_dir, overwrite = overwrite)
make_if_dont_exist(train_label_dir, overwrite = overwrite)
make_if_dont_exist(test_dir,overwrite= overwrite)
make_if_dont_exist(trained_model_dir, overwrite=overwrite)

It is best to have a construction like that now:

Picture by Creator

Setting the enironment variables

The script must know the place you place your raw_data, the place it might discover the preprocessed knowledge, and the place it needed to save the outcomes.

os.environ['nnUNet_raw_data_base'] = part of(main_dir,'nnUNet_raw_data_base')
os.environ['nnUNet_preprocessed'] = part of(main_dir,'preprocessed')
os.environ['RESULTS_FOLDER'] = trained_model_dir

Transfer the information in the correct repositories:

We outline a perform that can transfer our photographs to the correct repositories within the nnunet folder:

def copy_and_rename(old_location,old_file_name,new_location,new_filename,delete_original = False):
shutil.copy( part of(old_location,old_file_name),new_location)
os.rename( part of(new_location,old_file_name), part of(new_location,new_filename))
if delete_original:
os.take away( part of(old_location,old_file_name))

Now let’s run this perform for the enter and floor reality photographs:

list_of_all_files = os.listdir(segmentation_dir)
list_of_all_files = [file_name for file_name in list_of_all_files if file_name.endswith('.nii.gz')]

for file_name in list_of_all_files:

Now we have now to rename the information to be accepted by the nnUnet format, for instance topic.nii.gz will turn into subject_0000.nii.gz

def check_modality(filename):
examine for the existence of modality
return False if modality shouldn't be discovered else True
finish ='.nii.gz')
modality = filename[end-4:end]
for mod in modality:
if not(ord(mod)>=48 and ord(mod)<=57): #if not in 0 to 9 digits
return False
return True

def rename_for_single_modality(listing):

for file in os.listdir(listing):

if check_modality(file)==False:
new_name = file[:file.find('.nii.gz')]+"_0000.nii.gz"
os.rename( part of(listing,file), part of(listing,new_name))
print(f"Renamed to {new_name}")
print(f"Modality current: {file}")

# rename_for_single_modality(test_dir)

Organising the JSON file

We’re virtually completed!

You largely want to change 2 issues:

  1. The Modality (if its CT or MRI this modifications the normalization)
  2. The labels: Enter your personal lessons
overwrite_json_file = True #make it True if you wish to overwrite the dataset.json file in Task_folder
json_file_exist = False

if os.path.exists( part of(task_folder_name,'dataset.json')):
print('dataset.json exist already!')
json_file_exist = True

if json_file_exist==False or overwrite_json_file:

json_dict = OrderedDict()
json_dict['name'] = task_name
json_dict['description'] = "Segmentation of T1 Scans from MindBoggle"
json_dict['tensorImageSize'] = "3D"
json_dict['reference'] = "see problem web site"
json_dict['licence'] = "see problem web site"
json_dict['release'] = "0.0"

######################## MODIFY THIS ########################

#you might point out multiple modality
json_dict['modality'] = {
"0": "MRI"
#labels+1 must be talked about for all of the labels within the dataset
json_dict['labels'] = {
"0": "Non Mind",
"1": "Cortical grey matter",
"2": "Cortical White matter",
"3" : "Cerebellum grey ",
"4" : "Cerebellum white"


train_ids = os.listdir(train_label_dir)
test_ids = os.listdir(test_dir)
json_dict['numTraining'] = len(train_ids)
json_dict['numTest'] = len(test_ids)

#no modality in practice picture and labels in dataset.json
json_dict['training'] = [{'image': "./imagesTr/%s" % i, "label": "./labelsTr/%s" % i} for i in train_ids]

#eradicating the modality from check picture title to be saved in dataset.json
json_dict['test'] = ["./imagesTs/%s" % (i[:i.find("_0000")]+'.nii.gz') for i in test_ids]

with open( part of(task_folder_name,"dataset.json"), 'w') as f:
json.dump(json_dict, f, indent=4, sort_keys=True)

if os.path.exists( part of(task_folder_name,'dataset.json')):
if json_file_exist==False:
print('dataset.json created!')
print('dataset.json overwritten!')

Preprocess the info for nnU-Internet format

This creates the dataset for the nnU-Internet format

# -t 1 means "Task001", when you've got a special activity change it
!nnUNet_plan_and_preprocess -t 1 --verify_dataset_integrity

Practice the fashions

We are actually prepared to coach the fashions!

To coach the 3D U-Internet:

#practice 3D full decision U web
!nnUNet_train 3d_fullres nnUNetTrainerV2 1 0 --npz

To coach the 2D U-Internet:

# practice 2D U web
!nnUNet_train second nnUNetTrainerV2 1 0 --npz

To coach the cascade mannequin:

# practice 3D U-net cascade
!nnUNet_train 3d_lowres nnUNetTrainerV2CascadeFullRes 1 0 --npz
!nnUNet_train 3d_fullres nnUNetTrainerV2CascadeFullRes 1 0 --npz

Observe: If you happen to pause the traning and wish to resume it, add a “-c” ultimately for “proceed”.

For instance:

#practice 3D full decision U web
!nnUNet_train 3d_fullres nnUNetTrainerV2 1 0 --npz


Now we are able to run the inference:

result_dir = part of(task_folder_name, 'nnUNet_Prediction_Results')
make_if_dont_exist(result_dir, overwrite=True)

# -i is the enter folder
# -o is the place you wish to save the predictions
# -t 1 means activity 1, change it when you've got a special activity quantity
# Use -m second, or -m 3d_fullres, or -m 3d_cascade_fullres
!nnUNet_predict -i /content material/drive/MyDrive/neurosciences-segmentation/my_nnunet/nnUNet/nnunet/nnUNet_raw_data_base/nnUNet_raw_data/Task001/imagesTs -o /content material/drive/MyDrive/neurosciences-segmentation/my_nnunet/nnUNet/nnunet/nnUNet_raw_data_base/nnUNet_raw_data/Task001/nnUNet_Prediction_Results -t 1 -tr nnUNetTrainerV2 -m second -f 0 --num_threads_preprocessing 1

Visualization of the predictions

First let’s examine the coaching loss. This seems very wholesome, and we have now a Cube Rating > 0.9 (inexperienced curve).

That is really glorious for therefore little work and a 3D Neuroimaging segmentation activity.

7 Classes Realized on Making a Full Product Utilizing ChatGPT | by Shaked Zychlinski | Aug, 2023

Simulating Conductive Warmth Switch | In direction of Knowledge Science