Segmentation
A pixel-level classification task. In this project, each pixel is assigned to background, vertebrae, spinal canal, or discs.
Ahmed et al. 2025 / Modified U-Net
Four-Class Lumbar MRI Segmentation
This page summarizes a deep learning approach that classifies each MRI pixel into background, vertebrae, spinal canal, or intervertebral discs. The key idea is the combination of careful preprocessing, a modified U-Net, and a Combined Loss.
T2 SPACE segmentation workflow
MRI input
Modified U-Net softmax 4 classes
Predicted mask
Vertebrae Spinal Canal Intervertebral Discs
Paper PDF and Links
This section keeps the main paper, the ScienceDirect page, and the SPIDER dataset paper in one place for presentation preparation.
Main Paper PDF Open in new tab
Key Facts
The paper is important because it turns lumbar MRI segmentation into a cleaner four-class learning problem and reports very high performance on high-resolution T2 SPACE scans.
Dataset 218
Number of patients in the SPIDER lumbar MRI dataset.
Series 447
MRI series including T1, T2, and T2 SPACE sequences.
Classes 4
Background, vertebrae, spinal canal, and intervertebral discs.
Best Dice 0.97
Best reported performance on T2 SPACE images.
Data and Preprocessing
The original SPIDER data is stored as 3D MHA volumes. The paper converts it into 2D slices and merges detailed labels into four clinically useful classes.
1
Extract 2D slices from 3D MHA volumes
This reduces GPU memory cost and makes the data suitable for a 2D U-Net pipeline.
2
Merge detailed labels into four classes
Individual vertebra and disc IDs are converted into broader anatomical categories.
3
Filter imbalanced slices
Slices dominated by background or missing key structures are removed to stabilize training.
4
Compare T1, T2, and T2 SPACE
T2 SPACE achieves the best result because its higher resolution makes anatomical boundaries clearer.
SPIDER Label Mapping
0 Background
1-99 Vertebrae
100 Spinal Canal
200+ Intervertebral Discs
The key point is that the model predicts clinically useful anatomical categories, not individual anatomical IDs.
Model and Algorithm
The proposed model is based on U-Net. It compresses image features in the encoder and reconstructs a pixel-level segmentation map in the decoder.
Encoder
Convolution, Batch Normalization, Leaky ReLU, and Max Pooling extract increasingly abstract features.
Bottleneck
A 512-channel layer captures complex anatomical patterns and boundary information.
Skip Connection
Fine spatial information is passed from encoder to decoder to preserve boundaries.
Decoder
Transposed convolution restores resolution and reconstructs class-specific masks.
Softmax
Each pixel is assigned probabilities over the four output classes.
Combined Loss = 0.6 × Focal Loss + 0.4 × Dice Loss
Leaky ReLU
Keeps a small gradient for negative inputs and reduces the risk of inactive neurons.
Glorot Initialization
Stabilizes the starting weight distribution and helps gradients flow through deeper layers.
Focal + Dice
Balances hard-pixel learning with direct optimization of mask overlap.
Performance Metrics
Dice is the main metric. A value closer to 1 means stronger overlap between the predicted mask and the ground-truth annotation.
| 構造 | Dice | IoU | 意味 |
|---|---|---|---|
| Intervertebral Discs | 0.9688 | 0.9476 | Thin disc structures are segmented with high overlap. |
| Vertebrae | 0.9712 | 0.9461 | Large bony structures are segmented consistently. |
| Spinal Canal | 0.9671 | 0.9501 | The long canal-like structure remains accurate despite its shape. |
Dice
Measures overlap between prediction and ground truth. It is the easiest main metric to explain.
IoU
Intersection divided by union. It is usually stricter than Dice.
ASD / NSD
Boundary-distance metrics used to evaluate how far predicted surfaces deviate from the annotation.
Presentation note: Dice around 0.97 is very high, but the reported result should be discussed carefully because test-set details and preprocessing choices affect generalization.
Terminology
Use this section to quickly explain the technical terms during an English presentation.
U-Net
A widely used medical image segmentation architecture with an encoder, decoder, and skip connections.
T2 SPACE
A high-resolution 3D T2-weighted MRI sequence. It produces clearer anatomical boundaries.
Dice Coefficient
A measure of overlap between prediction and ground truth. Higher is better, with 1 meaning perfect overlap.
IoU
Intersection over Union. It divides the overlapping region by the total combined region.
Focal Loss
A loss function that gives more weight to hard-to-classify pixels.
Dice Loss
A loss function that directly optimizes the overlap between predicted and true masks.
Leaky ReLU
An activation function that keeps a small gradient for negative values.
Glorot / Xavier Initialization
A weight initialization method designed to keep training stable in deep networks.
Class Imbalance
A situation where some classes, such as background, dominate the image and can bias learning.
Graduation Project Plan
For the project, the first goal is reproduction. The second goal is improvement through model, loss, augmentation, and error-analysis experiments.
Reproduce the paper
Implement the SPIDER preprocessing pipeline, four-class labels, Modified U-Net, and Combined Loss.
Improve the method
Compare Attention U-Net, U-Net++, Boundary Loss, Tversky Loss, and stronger augmentation.
Visualize and analyze
Overlay predictions on MRI images and identify which structures or sequences fail most often.
Prepare the presentation
Explain the clinical motivation, technical method, reproduction result, limitations, and proposed improvements.