MONAI 101 tutorial - bioMONAI

Setup imports

import os
import tempfile

from bioMONAI.data import *
from bioMONAI.core import *
from bioMONAI.metrics import accuracy
from bioMONAI.losses import CrossEntropyLossFlat

from monai.networks.nets import DenseNet121
from monai.apps import MedNISTDataset

Setup data directory

You can specify a directory.
This allows you to save results and reuse downloads.
If not specified a temporary directory will be used.

base_directory = '../_data/'
if base_directory is not None:
    os.makedirs(base_directory, exist_ok=True)
root_dir = tempfile.mkdtemp() if base_directory is None else base_directory
print(root_dir)

../_data/

Use MONAI transforms to preprocess data

Medical images require specialized methods for I/O, preprocessing, and augmentation. They often follow specific formats, are handled with specific protocols, and the data arrays are often high-dimensional.

In this example, we will perform image loading, data format verification, and intensity scaling with three monai.transforms listed below, and compose a pipeline ready to be used in next steps.

from monai.transforms import LoadImageD, EnsureChannelFirstD, ScaleIntensityD, Compose

transform = Compose(
    [
        LoadImageD(keys="image", image_only=True),
        EnsureChannelFirstD(keys="image"),
        ScaleIntensityD(keys="image"),
    ]
)

Prepare datasets using MONAI Apps

We use MedNISTDataset in MONAI Apps to download a dataset to the specified directory and perform the pre-processing steps in the monai.transforms compose.

The MedNIST dataset was gathered from several sets from TCIA, the RSNA Bone Age Challenge, and the NIH Chest X-ray dataset.

The dataset is kindly made available by Dr. Bradley J. Erickson M.D., Ph.D. (Department of Radiology, Mayo Clinic) under the Creative Commons CC BY-SA 4.0 license.

If you use the MedNIST dataset, please acknowledge the source.

train_ds = MedNISTDataset(root_dir=root_dir, transform=transform, section="training", download=False)
val_ds = MedNISTDataset(root_dir=root_dir, transform=transform, section="validation", download=False)
test_ds = MedNISTDataset(root_dir=root_dir, transform=transform, section="test", download=False)

Loading dataset: 100%|██████████| 47164/47164 [00:28<00:00, 1632.66it/s]
Loading dataset: 100%|██████████| 5895/5895 [00:03<00:00, 1545.11it/s]
Loading dataset: 100%|██████████| 5895/5895 [00:03<00:00, 1692.22it/s]

data_ops = {
    'x_keys': 'image',
    'y_keys': 'label', 
    'bs': 512,
    'shuffle': True,
    'vocab': ['AbdomenCT','BreastMRI','ChestCT','CXR','Hand','HeadCT'],
    'show_summary': True,
}

data = BioDataLoaders.from_monai(train_ds, val_ds, **data_ops)

Train DataLoader
----------------
Dataset size : 47164
Batch size   : 512
Batches      : 93
Classes      : ['AbdomenCT', 'BreastMRI', 'ChestCT', 'CXR', 'Hand', 'HeadCT']

Batch structure:
  [0] shape=(512, 1, 64, 64) dtype=torch.float32 ~8.00 MB
  [1] shape=(512,) dtype=torch.int64 ~0.00 MB
Approx batch memory: 8.00 MB

Valid DataLoader
----------------
Dataset size : 5895
Batch size   : 1024
Batches      : 6
Classes      : ['AbdomenCT', 'BreastMRI', 'ChestCT', 'CXR', 'Hand', 'HeadCT']

Batch structure:
  [0] shape=(1024, 1, 64, 64) dtype=torch.float32 ~16.00 MB
  [1] shape=(1024,) dtype=torch.int64 ~0.01 MB
Approx batch memory: 16.01 MB

data.show_batch()

Define a network and a supervised trainer

To train a model that can perform the classification task, we will use the DenseNet-121 which is known for its performance on the ImageNet dataset.

For a typical supervised training workflow, MONAI provides SupervisedTrainer to define the hyper-parameters.

import torch 
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()

max_epochs = 5
model = DenseNet121(spatial_dims=2, in_channels=1, out_channels=6)
loss_function = CrossEntropyLossFlat()
metrics = [accuracy]

trainer = fastTrainer(data, model, loss_fn=loss_function, metrics=metrics, show_summary=False, lr=1e-5)

Run the training

trainer.fit(max_epochs)

epoch	train_loss	valid_loss	accuracy	time
0	0.697999	0.348656	0.969126	00:12
1	0.252962	0.123788	0.986768	00:12
2	0.109565	0.064564	0.990331	00:12
3	0.055369	0.037543	0.994402	00:11
4	0.032297	0.023173	0.997286	00:11

torch.cuda.max_memory_allocated() / 1024**2

5405.203125

trainer.show_results()

Check the prediction on the test dataset

test_dl = test_biodataloader(data, test_ds) # type:ignore

evaluate_classification_model(trainer, test_data=test_dl, metrics=accuracy, show_graph=True, show_results=True); # type:ignore

              precision    recall  f1-score   support

   AbdomenCT       0.99      1.00      1.00       993
   BreastMRI       1.00      1.00      1.00       903
         CXR       1.00      0.99      1.00       968
     ChestCT       1.00      1.00      1.00       998
        Hand       0.99      1.00      1.00       996
      HeadCT       1.00      1.00      1.00      1037

    accuracy                           1.00      5895
   macro avg       1.00      1.00      1.00      5895
weighted avg       1.00      1.00      1.00      5895

Most Confused Classes:

	Actual Class	Predicted Class	Count
0	CXR	Hand	5
1	HeadCT	AbdomenCT	3
2	BreastMRI	AbdomenCT	2
3	ChestCT	AbdomenCT	1
4	Hand	CXR	1
5	Hand	HeadCT	1

	Value
CrossEntropyLossFlat
Mean	1.057990
Median	1.049570
Standard Deviation	0.044332
Min	1.044239
Max	1.902058
Q1	1.047402
Q3	1.054644

	Value
accuracy
Mean	0.997795
Median	1.000000
Standard Deviation	0.046908
Min	0.000000
Max	1.000000
Q1	1.000000
Q3	1.000000