niftiai

niftiai aims to be the easiest framework to train neural nets on 3D images (often NIfTIs), using

• fastai easy neural net training
• niftiview easy 3D image viewing
• mriaug easy 3D image (+MRI-specific) augmentation

pip install niftiai to simplify your code and skip complex frameworks like MONAI and torchio!

Quick Start 🚀

Study the beginner tutorial to understand how 15 lines of code can train a neural net to classify 🧠 MRIs...

import openneuro as on
from deepmriprep import run_preprocess
from fastai.basics import pd, accuracy, CategoryBlock
from niftiai import cnn_learner3d, Scale, ImageDataLoaders3d
DATA_DIR = 'data/ds000005'

on.download(dataset=DATA_DIR[-8:], target_dir=DATA_DIR, include='*/anat/*T1w.*')
df = run_preprocess(bids_dir=DATA_DIR, outputs=['brain'])
table = pd.read_table(f'{DATA_DIR}/participants.tsv')
df = pd.concat([table, df.reset_index()], axis=1)
dls = ImageDataLoaders3d.from_df(df, fn_col='brain', item_tfms=[Scale()],
                                 label_col='sex', y_block=CategoryBlock())
learner = cnn_learner3d(dls, metrics=[accuracy])
learner.fit(3)
learner.show_results()  # add plt.show() if not in Jupyter notebook

...and the intermediate tutorial to see how 12 lines of code train a UNet to do MR image segmentation 🤯

import openneuro as on
from deepmriprep import run_preprocess
from niftiai import unet_learner3d, aug_transforms3d, Scale, SegmentationDataLoaders3d
DATA_DIR = 'data/ds000001'

on.download(dataset=DATA_DIR[-8:], target_dir=DATA_DIR, include='*/anat/*T1w.*')
df = run_preprocess(bids_dir=DATA_DIR, outputs=['mask']).reset_index()
aug_tfms = aug_transforms3d()
dls = SegmentationDataLoaders3d.from_df(df, fn_col='t1', item_tfms=Scale(),
                                        label_col='mask', batch_tfms=aug_tfms)
learner = unet_learner3d(dls, c_out=2, linknet=True)
learner.fit(3, lr=1e-2)
learner.show_results()  # add plt.show() if not in Jupyter notebook

👩‍💻 Also study the fastai tutorials to understand the underlying framework that is

... approachable and rapidly productive, while also being deeply hackable...

and discover its wide range of features like

• mixed precision training for reduced VRAM usage via Learner.to_fp16() and Learner.to_bf16()

learner = learner.to_fp16()  # to enable FP16, use this line before training
learner = learner.to_bf16()  # to enable BF16, use this line before training

• 1cycle scheduled training for faster convergence via Learner.fit_one_cycle()

from fastai.callback.all import *
...
learner.fit_one_cycle(...)  # instead of learner.fit(...) to enable 1cycle scheduled training

• distributed training to use multiple GPUs via accelerate

from fastai.distributed import *
...
with learner.distrib_ctx(): learn.fit(...) # and run with "accelerate launch ..."

• ...so much more...