Deep Learning · Medical Imaging

Brain Tumor
Segmentation

Automated tumor detection and segmentation from MRI scans using a U-Net convolutional architecture trained on brain tumor datasets.

Try Demo View Pipeline

U-NetArchitecture

256²Input Resolution

5Training Epochs

BCELoss Function

Interactive Demo

Upload an MRI Scan

Upload a brain MRI image to visualize the segmentation process. The demo shows how the model identifies and highlights tumor regions.

🧠

Click or drag & drop PNG, JPG, WEBP supported

◉ Awaiting input
─────────────
Upload MRI to begin

Methodology

Processing Pipeline

Data Acquisition

MRI brain scans sourced from Kaggle's Brain Tumor Segmentation dataset. Images include paired grayscale masks marking tumor boundaries.

Kaggle Dataset

Pre-processing

Images are loaded, cropped to region of interest (rows 90–450, cols 150–406), resized to 256×256, and normalized to [0,1]. Masks converted to binary boolean arrays.

NumPy · TensorFlow

U-Net Training

5-block encoder–decoder with skip connections. Adam optimizer, binary cross-entropy loss. EarlyStopping with patience=2 prevents overfitting. 10% validation split.

Keras · EarlyStopping

Prediction & Visualization

Model outputs a sigmoid probability mask. Predictions visualized as standalone mask and alpha-blended overlay on the original scan.

Matplotlib · cv2

Model Design

U-Net Architecture

Encoder–decoder with skip connections enabling precise spatial localization.

📉

Encoder (Contraction)

5 levels: 16→32→64→128→256 filters. Each block has 2× Conv2D (3×3, ReLU) + Dropout + MaxPooling. Progressively captures abstract features.

📈

Decoder (Expansion)

Conv2DTranspose upsamples feature maps. Concatenated with encoder skip connections to recover fine spatial details lost during pooling.

🔗

Skip Connections

Direct connections from encoder to decoder at each level. Ensures high-resolution spatial information is preserved for precise segmentation boundaries.

🎯

Output Layer

1×1 Conv2D with sigmoid activation produces a single-channel probability map. Threshold at 0.5 for binary tumor/non-tumor classification.

Python · TensorFlow

# U-Net Output Head
outputs = tf.keras.layers.Conv2D(1, (1, 1), activation='sigmoid')(c9)
model   = tf.keras.Model(inputs=[inputs], outputs=[outputs])
model.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy']
)

# Training with EarlyStopping
history = model.fit(
    X_train, y_train,
    validation_split=0.1,
    batch_size=32,
    epochs=5,
    callbacks=[EarlyStopping(monitor='val_loss', patience=2)]
)

Brain TumorSegmentation