Multi-class semantic segmentation

For the multi-class semantic segmentation task, we will use the brain tumors dataset from the Medical Segmentation Decathlon challenge (http://medicaldecathlon.com/). The data is collected from the Multimodal Brain Tumor Image Segmentation Benchmark Challenge (BraTS) dataset from 2016 and 2017. The task is to segment tumors into three different subregions (active tumor (AT), necrotic core (NCR), and peritumoral edematous/infiltrated tissue (ED)) from multimodal multisite MRI data (T1w, T1ce, T2w, and FLAIR). The challenge with this dataset is the brain tumors’ highly heterogeneous appearance and shape.

Google Colab

from fastMONAI.vision_all import *

from monai.apps import DecathlonDataset
from sklearn.model_selection import train_test_split

Download external data

We use the MONAI function DecathlonDataset to download the data and generate items for training.

path = Path('../data')
path.mkdir(exist_ok=True)

task = "Task01_BrainTumour"
training_data = DecathlonDataset(root_dir=path, task=task, section="training", download=True,
                                 cache_num=0, num_workers=3)

2025-08-29 09:44:58,814 - INFO - Verified 'Task01_BrainTumour.tar', md5: 240a19d752f0d9e9101544901065d872.
2025-08-29 09:44:58,815 - INFO - File exists: ../data/Task01_BrainTumour.tar, skipped downloading.
2025-08-29 09:44:58,816 - INFO - Non-empty folder exists in ../data/Task01_BrainTumour, skipped extracting.

df = pd.DataFrame(training_data.data)
df.shape

(388, 2)

Split the labled data into training and test

train_df, test_df = train_test_split(df, test_size=0.1, random_state=42)
train_df.shape, test_df.shape

((349, 2), (39, 2))

Look at training data

med_dataset = MedDataset(img_list=train_df.label.tolist(), dtype=MedMask, max_workers=12)

med_dataset.df.head()

	path	dim_0	dim_1	dim_2	voxel_0	voxel_1	voxel_2	orientation	voxel_count_0	voxel_count_1	voxel_count_2	voxel_count_3
0	../data/Task01_BrainTumour/labelsTr/BRATS_477.nii.gz	240	240	155	1.0	1.0	1.0	RAS+	8765377	83088	15826	63709.0
1	../data/Task01_BrainTumour/labelsTr/BRATS_350.nii.gz	240	240	155	1.0	1.0	1.0	RAS+	8872636	21364	8872	25128.0
2	../data/Task01_BrainTumour/labelsTr/BRATS_266.nii.gz	240	240	155	1.0	1.0	1.0	RAS+	8725071	83276	69784	49869.0
3	../data/Task01_BrainTumour/labelsTr/BRATS_294.nii.gz	240	240	155	1.0	1.0	1.0	RAS+	8790699	90806	20231	26264.0
4	../data/Task01_BrainTumour/labelsTr/BRATS_466.nii.gz	240	240	155	1.0	1.0	1.0	RAS+	8911252	14046	60	2642.0

summary_df = med_dataset.summary()

summary_df.head()

	dim_0	dim_1	dim_2	voxel_0	voxel_1	voxel_2	orientation	example_path	total
0	240	240	155	1.0	1.0	1.0	RAS+	../data/Task01_BrainTumour/labelsTr/BRATS_002.nii.gz	349

suggestion = med_dataset.get_suggestion()
target_spacing = suggestion['target_spacing']
target_spacing

# Display dataset size statistics
stats = med_dataset.get_size_statistics(target_spacing=target_spacing)
print(f"Image sizes (after resampling to {target_spacing}):")
print(f"  Min:    {stats['min']}")
print(f"  Median: {stats['median']}")
print(f"  Max:    {stats['max']}")

# Use max size for padding (ensures all images fit)
img_size = stats['max']
print(f"\nTarget size (max): {img_size}")

bs=4
# Manual size override - larger images will be cropped to fit
size = [224, 224, 128]

item_tfms = [ZNormalization(), PadOrCrop(size), RandomAffine(scales=0, degrees=5, isotropic=True)]

dblock = MedDataBlock(blocks=(ImageBlock(cls=MedImage), MedMaskBlock), 
                      splitter=RandomSplitter(seed=42),
                      get_x=ColReader('image'),
                      get_y=ColReader('label'),
                      item_tfms=item_tfms,
                      apply_reorder=True,  # Reorder to RAS+ orientation (the common case)
                      target_spacing=target_spacing)

dls = dblock.dataloaders(train_df, bs=bs)

# training and validation
len(dls.train_ds.items), len(dls.valid_ds.items)

(280, 69)

dls.show_batch(anatomical_plane=0)

Create and train a 3D model

As in the binary segmentation task, we import an enhanced version of UNet from MONAI. This time instead of using Dice loss, we import a loss function that combines Dice loss and Cross Entropy loss and returns the weighted sum of these two losses.

from monai.losses import DiceCELoss
from monai.networks.nets import UNet

codes = np.unique(med_img_reader(train_df.label.tolist()[0]))
n_classes = len(codes)
print("Unique classes:",*codes)

Unique classes: 0.0 1.0 2.0 3.0

monai_model = UNet(spatial_dims=3, in_channels=4, out_channels=n_classes, channels=(16, 32, 64, 128, 256),strides=(2, 2, 2, 2), num_res_units=2)
pytorch_model = monai_model.model

loss_func = CustomLoss(loss_func=DiceCELoss(to_onehot_y=True, include_background=True, softmax=True))

learn = Learner(dls, monai_model, loss_func=loss_func, opt_func=ranger, metrics=multi_dice_score)#.to_fp16()

learn.lr_find()

SuggestedLRs(valley=0.001737800776027143)

lr = 1e-1

from fastMONAI.utils import create_mlflow_callback

mlflow_callback = create_mlflow_callback(learn, experiment_name=task)
learn.fit_flat_cos(2, lr, cbs=[mlflow_callback])

learn.save('braintumor-weights')

Path('models/braintumor-model.pth')

learn.show_results(anatomical_plane=0, ds_idx=1)

mlflow_ui = MLflowUIManager()
mlflow_ui.start_ui()

Inference on test data

learn.load('braintumor-weights');

test_dl = learn.dls.test_dl(test_df[:10],with_labels=True)

test_dl.show_batch(anatomical_plane=0, figsize=(10,10))

pred_acts, labels = learn.get_preds(dl=test_dl)
pred_acts.shape, labels.shape

(torch.Size([10, 4, 224, 224, 128]), torch.Size([10, 1, 224, 224, 128]))

Dice score for labels 1,2 and 3:

multi_dice_score(pred_acts, labels)

tensor([0.5708, 0.4186, 0.6994])

learn.show_results(anatomical_plane=0, dl=test_dl)

mlflow_ui.stop()