This project focuses on segmenting brain tumors from MRI scans using a manually implemented U-Net architecture in TensorFlow. The model is trained and tested on preprocessed brain images with corresponding ground truth masks.
- Manually implemented U-Net architecture in TensorFlow
- End-to-end pipeline: data loading, preprocessing, training, evaluation
- Segmentation performance visualized using original images, ground truth masks, and predicted masks
- Works with binary tumor masks and grayscale brain MRIs
- Model tested on real-world cases to validate prediction accuracy
Visual examples of the model's performance on unseen data.
The model uses a custom-built U-Net, designed from scratch in TensorFlow.
Input (256x256x1)
|
[E1]─┬→ [P1] →
| |
[E2]─┬→ [P2] →
| |
[E3]─┬→ [P3] →
| |
[E4]─┬→ [P4] →
| |
[Bottleneck (B)]
|
[D1] ←─── skip: [E4]
|
[D2] ←─── skip: [E3]
|
[D3] ←─── skip: [E2]
|
[D4] ←─── skip: [E1]
|
Output (256x256x1)
[E1]to[E4]: Encoder blocks (encoder_block)[P1]to[P4]: Pooling layers (MaxPooling2D)[B]: Bottleneck (conv_block)[D1]to[D4]: Decoder blocks (decoder_block)- Arrows
←─── skip:represent skip connections via concatenation - Final
Output: Single-channel (Conv2D(1, (1,1), activation='sigmoid'))
Organized into separate folders for input images and masks:
datasets/
└── Brain-Tumor-Classification-2D/
└── Astrocytoma/
git clone https://github.com/haelahpx/BrainTumorSegmentation.git
cd brain-tumor-segmentation├── .git/ # Git configuration folder
├── assets/ # Images and visual assets (e.g., model diagrams)
├── datasets/ # Dataset files (e.g., training and testing data)
├── models/ # Saved models or model checkpoints
├── .gitattributes # Git attributes configuration
├── LICENSE # License information
├── Mini-Project # Report or project description document
├── README # Project overview (you can rename to README.md)
├── testing # Test scripts or logs
- Python
- TensorFlow / Keras
- NumPy, Matplotlib
This is a private research/learning project. Contact for permissions or collaborations.