1. Department of Computer Sc. & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree
Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh)
Sector-26, Chandigarh - 160019
SUMMER TRAINING
4th SEMESTER
SKULL RECONSTRUCTION
Presenter – Ishtveer Singh Billing (CO21325)
2. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Topics We’ll Cover
• Why Skull Reconstruction
• Database Creation/ Selection
• Data Preprocessing
• Implementing U-Net Architecture
• Encoder & Decoder
• 3D Unet Model
• Training the Model
• Future Work
CS-604
3. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Why Skull Reconstruction
Skull provides basic information needed to determine the sex, stage in
life, and ethnic origin of the deceased.
2D imaging offers limited visualization in terms of depth and spatial
relationships.
It allows surgeons to simulate surgery procedures, assess the feasibility
of different approaches, and optimize surgical outcomes.
It helps forensic facial reconstruction.
Skull reconstruction in 3D form allows for a more comprehensive and
detailed understanding of the skull's shape, structure, and anatomical
features.
4. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Obtain a dataset of skull images with corresponding ground truth labels.
These images can be obtained from medical imaging databases or collected
through other means.
A good dataset to work with has the following properties:
Representativeness: The dataset should be representative of the real-world
problem covering wide range of possible cases.
Balance: There should be an equal number of examples of each class.
Size: The dataset should be large enough to train the model effectively. .
Quality: The data should be of high quality and should not contain any
errors or outliers. .
Scalability: It should be possible to add new data to the dataset without
having to retrain the model.
A Good Dataset?
5. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
6. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Data Pre-processing
Converting .nrrd to .nii files
7. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
8. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
9. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Changing Size
Here are some of the steps that can be done during preprocessing:
RESIZING: It can help to improve the performance of the model
because the model will learn to segment the skull regardless of its
size.
NORMALISING: Normalizing the pixel values can help to remove
the noise from the images and make the pixel values more
consistent.
SEGMENTATION: Segmentation can be done before or after
implementation. It helps to improve the performance of the
model
10. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
11. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
12. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Masking Skull Image
13. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
14. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Implementing U-Net Architechture
15. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Encoder and Decoder
The U-Net architecture consists of two parts: an encoder and a
decoder. The encoder is responsible for extracting features from the
input image. It does this by using a series of convolutional layers,
which are followed by max pooling layers. The max pooling layers
reduce the size of the feature maps, which helps to prevent overfitting.
The decoder is responsible for reconstructing the output image from
the features extracted by the encoder. It does this by using a series of
convolutional layers, which are followed by upsampling layers. The
upsampling layers increase the size of the feature maps, which helps to
reconstruct the output image.
16. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Segmenting medical images:
U-Net has been used to segment medical images, such as MRI
scans and CT scans. This can be used to identify different
organs and tissues in the body.
Segmenting natural images:
U-Net has also been used to segment natural images, such as
images of cells and images of flowers. This can be used to
identify different objects in the image.
Segmenting remote sensing images:
U-Net has also been used to segment remote sensing images,
such as images of urban areas and images of forests. This can
be used to identify different land cover types in the image.
Benefits of U-Net Model
17. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
18. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Defining 3D Unet Model
19. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Training the Model
Optimizer:
It is responsible for updating the model's weights during training. A good
optimizer for U-Net models is Adam, which is a stochastic gradient descent
(SGD) optimizer with adaptive learning rates.
Loss function:
The loss function is used to measure the difference between the model's
predictions and the ground truth labels. A good loss function for U-Net models is
binary cross-entropy, which is a loss function that is specifically designed for
binary classification problems.
Metrics:
Metrics are used to evaluate the performance of the model during training and
testing. Some good metrics for U-Net models include accuracy, precision, and
recall.
Training Parameters
20. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Epochs:
The number of epochs is the number of times that the model will go through
the entire training dataset. A good number of epochs for U-Net models is
typically between 100 and 200.
Batch size:
The batch size is the number of images that will be processed at a time during
training. A good batch size for U-Net models is typically between 16 and 32.
Data augmentation:
Data augmentation can be used to increase the size of the training dataset and
to make the model more robust to variations in the data. Some common data
augmentation techniques for U-Net models include random cropping,
random flipping, and random rotation.
Learning rate:
The learning rate is the rate at which the model's weights are updated during
training. A good learning rate for U-Net models is typically between 0.001 and
0.0001.
Training Parameters
21. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
22. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
23. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
Future Work
• Use a larger dataset: The current dataset is relatively small. Using a larger
dataset would help to improve the performance of the model.
• Use a different loss function: The current loss function is binary cross-
entropy. Using a different loss function, such as Dice loss, could improve the
performance of the model.
• Use a different optimizer: The current optimizer is Adam. Using a different
optimizer, such as SGD with momentum, could improve the performance of
the model.
• Use a different architecture: The current architecture is U-Net. Using a
different architecture, such as DeepLabv3+, could improve the performance of
the model.
• Use a different modality: The current modality is RGB images. Using a
different modality, such as CT scans or MRI scans, could improve the
performance of the model.
• Deploy the model: Once the model is trained, it can be deployed to a web
application or a mobile app. This would allow users to use the model to
segment skulls in real time.
24. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
References
1. github.com/RifugioTorino/my
2. github.com/itakuto/AppleTomatoWeb
3. en.wikipedia.org/wiki/Image_segmentation
4. en.wikipedia.org/wiki/U-Net
5. MUG500+ Repository (figshare.com)
25. Department of Computer Science & Engineering
Chandigarh College of Engineering & Technology (CCET -Degree Wing)
(A Govt. College under Chandigarh UT Administration, Chandigarh) ,Sector-26, Chandigarh - 160019
THANK YOU!
Any Queries?