📌 Overview

This project implements a U-Net-based convolutional neural network for automatic segmentation of spinal structures from MRI scans. U-Net, originally designed for biomedical image segmentation, is particularly well-suited for this task due to its ability to capture both spatial context and fine-grained details.

🏗️ U-Net Architecture

The architecture consists of three main components:

Encoder (Contracting Path)

Extracts hierarchical features while progressively reducing spatial resolution using:

• Two 3x3 convolutions with ReLU activation
• 2x2 max pooling (stride 2) for downsampling
• Five encoder levels in total

Bridge (Bottleneck)

• Connects encoder and decoder paths
• Captures deep, abstract representations of features

Decoder (Expanding Path)

• Mirrors the encoder structure (five levels)
• 2x2 up-convolution (transposed convolution) layers for upsampling
• Skip connections concatenate encoder feature maps with upsampled features
• Two 3x3 convolution layers with ReLU at each level
• Final output layer with Softmax activation to generate segmentation masks

⚙️ Technical Implementation

• Input: Grayscale MRI images (1 channel)
• Output: Binary mask (intervertebral disc vs. background)
• Padding: "Same" padding to preserve spatial dimensions
• ReLU Activation: ReLU(x) = max(0, x)
• Upsampling Rule: Output Size = Input Size × Stride
• Pooling Rule: Output = (Input - Pool Size) / Stride + 1

🧪 Training & Results

• Epochs: 100
• Batch Size: 8
• Total Training Time: 35 hours
• Output Format: H5 model file

After training, the model was evaluated on unseen MRI scans. Predicted masks showed strong alignment with ground truth labels, demonstrating effective generalization.

Visualization: Below is an example of the predicted segmentation masks (not shown here — insert image if available).

🔧 Tools & Libraries

• Python
• TensorFlow / Keras
• NumPy
• Matplotlib (for result visualization)

💻 U-Net Model Code (Simplified)

from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, concatenate
from tensorflow.keras.models import Model

inputs = Input((128, 128, 1))

# Encoder
c1 = Conv2D(64, (3, 3), activation='relu', padding='same')(inputs)
c1 = Conv2D(64, (3, 3), activation='relu', padding='same')(c1)
p1 = MaxPooling2D((2, 2))(c1)

# Bottleneck
c2 = Conv2D(128, (3, 3), activation='relu', padding='same')(p1)
c2 = Conv2D(128, (3, 3), activation='relu', padding='same')(c2)

# Decoder
u1 = UpSampling2D((2, 2))(c2)
m1 = concatenate([u1, c1])
c3 = Conv2D(64, (3, 3), activation='relu', padding='same')(m1)
c3 = Conv2D(64, (3, 3), activation='relu', padding='same')(c3)

outputs = Conv2D(1, (1, 1), activation='sigmoid')(c3)
model = Model(inputs, outputs)
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

📝 Conclusion

This spinal segmentation project highlights the power of U-Net for medical image analysis. The symmetrical architecture, use of skip connections, and strong spatial feature preservation enable precise segmentation of spinal discs from complex MRI data.