ditt-a/3D_UnetPlusPlus_Brain_Tumour_Segmentation

Output

The model predicts a 3-channel segmentation mask:

Channel 0: ET - Enhancing Tumor
Channel 1: TC - Tumor Core
Channel 2: WT - Whole Tumor

Intended Use

This model is intended for:

• Research in medical image segmentation
• Educational experiments with U-Net and U-Net++ architectures
• Brain tumor segmentation benchmarking
• Experiment comparison using Dice score metrics

Out-of-Scope Use

This model is not intended for direct clinical diagnosis or treatment decision-making. Predictions should not be used as a substitute for expert medical evaluation by radiologists or clinicians.

Training Data

The experiments are based on BraTS20-style brain MRI data. The original GitHub repository references the BRATS20 Training and Validation Dataset from Kaggle.

The dataset consists of multi-modal MRI volumes and segmentation masks in NIfTI format.

Preprocessing

The training notebooks use a medical imaging preprocessing pipeline that includes:

• Loading NIfTI MRI volumes
• Ensuring channel-first format
• Concatenating MRI modalities into a multi-channel input
• Intensity normalization
• Foreground cropping
• Spatial cropping with ROI size of 128 x 128 x 128
• Random intensity scaling
• Random intensity shifting
• Random flipping augmentation

The segmentation labels are converted into three tumor-region masks:

ET = label 4
TC = labels 1 or 4
WT = labels 1, 2, or 4

Training Procedure

The repository includes separate notebooks for training and inference:

intern_unet.ipynb     : U-Net experiment pipeline
intern_unetpp.ipynb   : U-Net++ experiment pipeline
prediksi.ipynb        : Prediction / inference workflow

The original experiment setup used an 80:20 train-validation split. In the available training log, 368 samples were found, with 294 samples used for training and 74 samples used for validation.

Evaluation

The main evaluation metric is Dice score.

Reported experiment outputs are stored in CSV files in the ReportHasil/ folder.

Model / Strategy	Best Validation Mean Dice	Notes
U-Net Raw	0.865	Best reported U-Net raw experiment
U-Net Warmup + Freeze	0.857	Warmup and encoder-freezing strategy
U-Net Warmup + LR	0.849	Warmup with learning-rate strategy
U-Net++ Raw	0.855	Baseline U-Net++ experiment
U-Net++ Warmup + Freeze	0.853	Warmup and freezing strategy
U-Net++ Warmup + LR	0.863	Best reported U-Net++ variant

The Dice scores are reported for:

• Mean Dice
• ET Dice
• TC Dice
• WT Dice

Example Inference Workflow

A typical inference workflow is:

1. Prepare the MRI modality files for one patient.
2. Load the trained checkpoint.
3. Apply the same preprocessing used during training.
4. Run sliding-window inference.
5. Convert the predicted 3-channel output into ET, TC, and WT masks.
6. Save or visualize the segmentation result.

Example expected patient folder structure:

Patient_ID/
├── Patient_ID_flair.nii.gz
├── Patient_ID_t1.nii.gz
├── Patient_ID_t1ce.nii.gz
├── Patient_ID_t2.nii.gz
└── Patient_ID_seg.nii.gz

Limitations

• The model was trained and evaluated on BraTS20-style MRI data, so performance may decrease on data from different scanners, hospitals, preprocessing pipelines, or patient populations.
• The model expects all required MRI modalities to be available.
• The model is designed for research use and has not been clinically validated.
• The reported metrics come from the project experiments and should be interpreted as validation results, not real-world clinical performance.

Ethical Considerations

Medical AI models can produce incorrect predictions, especially when used outside their training distribution. This model should be used only for research, education, or experimentation. Any medical interpretation must be reviewed by qualified healthcare professionals.

Repository

Original GitHub repository:

https://github.com/Ditt-A/BrainTumourSegmentation

License

This project is released under the MIT License.