This repository presents OncoWaveNet, a novel medical image segmentation framework that combines:
- Residual U-Net architecture
- Wavelet-based denoising
- Tumor-centric slice sampling
- Hybrid loss optimization
Trained on the BraTS 2020 Dataset, the model focuses on accurate segmentation of:
- 🟥 Enhancing Tumor (ET)
- 🟨 Tumor Core (TC)
- 🟩 Whole Tumor (WT)
- 🎯 Automate segmentation of brain tumor subregions in multimodal MRI scans.
- 🧼 Reduce noise and preserve structural features using wavelet denoising.
- 📐 Use a hybrid loss (Dice + Categorical Crossentropy) to improve boundary accuracy.
- 🧠 Enable efficient training under constrained hardware.
- ✅ Residual U-Net (ResUNet) backbone for deep feature learning.
- ✅ Wavelet Denoising using Daubechies-1 (Haar) filters for noise suppression.
- ✅ Z-axis slice sampling centered on tumor for high focus and reduced data redundancy.
- ✅ Class remapping + One-hot encoding for multi-class segmentation.
- ✅ Fully compatible with BraTS 2020 standardized annotations.
| Category | Tools / Libraries |
|---|---|
| Language | Python 3.8+ |
| Deep Learning | PyTorch, TensorFlow, Keras |
| Medical Imaging | NiBabel, SimpleITK, OpenCV |
| Data Handling | NumPy, Pandas, Matplotlib, Seaborn |
| Optimization | Adam Optimizer, ReduceLROnPlateau |
| Evaluation | Dice Coefficient, IoU, Accuracy, Precision, etc |
| Development | Jupyter Notebook, Google Colab, VS Code |
| Version Control | Git, GitHub |
Implemented using
SimpleITK,PyWavelets,NumPy, and custom utilities.
-
Tumor Slice Extraction:
Extract 100 axial slices centered around tumor Z-location for focused training. -
Wavelet Denoising:
Apply Haar wavelet transform to FLAIR & T1ce slices to suppress background noise. -
Z-score Normalization:
Normalize pixel intensities to zero-mean, unit variance per slice. -
Label Remapping & Encoding:
Convert label4 → 3and apply one-hot encoding to segment ET, TC, WT. -
Resizing:
Resize all slices to uniform spatial dimensions (e.g., 240×240).
Residual U-Net combines the U-Net encoder-decoder structure with residual blocks to improve gradient flow and feature reuse.
Skip connections preserve spatial resolution and fine features crucial for medical segmentation.
📌 Block Overview:
- Encoder: 4 residual downsampling blocks (C1–C4)
- Bottleneck: Deepest abstract features (C5)
- Decoder: 4 residual upsampling blocks with skip connections (C6–C9)
- Output: 1×1 convolution with softmax to predict segmentation classes
Residual Block = (Conv → BN → ReLU → Conv → BN) + Shortcut
- 369 patients: 259 High Grade Gliomas (HGG), 110 Low Grade (LGG)
- MRI modalities: T1, T1ce, T2, FLAIR
- Voxel-wise labels: 0 = Background, 1 = Necrosis, 2 = Edema, 4 = Enhancing Tumor
- All images registered to 240×240×155 with 1mm³ resolution
| Metric | ET | TC | WT |
|---|---|---|---|
| Dice | 0.7779 | 0.8789 | 0.8920 |
| Overall Performance | Value |
|---|---|
| Accuracy | 99.58% |
| Precision | 99.58% |
| Sensitivity | 99.58% |
| Specificity | 99.86% |
📉 See results/metrics.png and results/graphs/ for training curves and evaluation plots.
- 📈 Extend to 3D ResUNet or 2.5D hybrid models for volumetric consistency
- 🧩 Add attention mechanisms (e.g., SE blocks, Transformers)
- 🎨 Visualize segmentation overlays with ITK-SNAP or 3D Slicer
- 🧬 Integrate Generative AI (LDM) for tumor progression prediction
- 📝 Use LLMs (GPT-4) + Jinja2 for automatic radiology report generation
-
Shaikh Asif Hossain
B.Tech, Calcutta Institute of Engineering & Management
📧 hossainshaikhasif@gmail.com
🌐 GitHub -
Surojit Bose
B.Tech, Calcutta Institute of Engineering & Management 📧 surojitbose2002@gmail.com -
Debam Saha
Assistant Professor, Department of CSE, CIEM 📧 debamsaha@ciem.co.in
This repository is licensed under the MIT License.
Feel free to use, adapt, or extend with attribution.
If you find this project helpful, please ⭐ star the repository and share it with your peers!