Student Experience

STFC/Durham University CDT in Data Intensive Science.

Carolina Cuesta-Lázaro

Arnau Quera-Bofarull

(Joseph Bullock)

Placement at IBEX Innovations Ltd.

Who are we?

2 months team project at IBEX innovations

Carolina / Arnau

Cosmology

Joe

Particle Physics

Detect bone and soft-tissue on X-Ray images

Detect

collimator

Segment

Open beam

Bone

Soft-tissue

Previous approaches

• Challenging problem due to varying brightness throughout the image.
• Usually done by detecting edges and shapes.
• High accuracy requires tuning of hyper parameters per image and body-part (not automated).
• Not well defined boundaries.

 Kazeminia, S., Karimi, et al (2015)

Are there better features hidden in the data?

Extracts features from a high dimensional feature space, once trained on a particular dataset.

DEEP LEARNING

What is training?

Luminosity

Size

Colour

Galaxy

Star

Deep learning on images

Credit : https://www.pnas.org/content/116/4/1074

With the right data, the network will find the right features

With wrong data ....

Does it really work?

SegNet

• The network has more than 15 Million free parameters.
• To find the values of the parameters that produce the correct segmentation, it has been trained on 1.3 Million images.

Could it solve our problem?

150 labeled images.

Hardware limitations (memory, training time...).

We need a fast network, easier to re-train as we get more images.

CONS

The Road to XNet

Coursera

The Dataset

• Small, ~150 images
• Unbalanced

Solution:

Artificially augment the dataset by transforming original images.

Splitting the dataset

We typically divide our dataset into three subsets:

1. Training: 70% from categories with more than one sample.
   Data augmentation -> Equal sample size for all categories.
   Final size ~ 7000 images.
2. Validation: 15%  from categories with more than one sample. Used to stop training and hyperparameter tuning.
3. Test: 15%  including categories with only one sample. Final network performance is evaluated in this set.

Network Architecture

First attempts focused on a very simplified SegNet model.

Network Architecture

Going deeper has limits ( limited image size, GPU memory bottleneck, overfitting).

Dealing with overfitting

Ways to reduce overfitting

1. Increase the dataset.
2. Reduce network complexity.
3. Regularisation.

Credit: www.kdnuggets.com

XNet

1. Typical encoder-decoder architecture.
2. W-shape for two feature extraction stages. Avoids resolution problems.
3. Skip connections across levels.
4. L2 regularisation at each convolutional layer.

Final model: XNet

Post-processing

The network outputs 3 probability maps.

Soft tissue probability map

Results

• Generalises well even for unseen categories!
• Overall accuracy on test set: 92%
• Soft tissue TP/FP rate: 82% / 4%

Comparison with other methods

• Smooth connected boundaries.
• Better generalisation to different body parts (we do not have any frontal view of a foot in our dataset).
• More robust to noise.
• Well defined metrics to benchmark against.

• The development process takes a long time due to hyperparameter tunning (50% of our internship time).
• We used ~1000 GPU hours.
  • 3x 4GB GPUS
  • 1x 8GB GPU
  • AWS 12GB GPU

Development process

Cryptocurrency Times

• Promising ML applications to medical imaging.
• Possible to train ML models with limited hardware and resources.

• Knowledge of building and deploying a machine learning product in an industrial setting.

• Paper out arXiv:1812.00548v1, and presented at the SPIE Medical Imaging conference in San Diego. Best student paper awarded.

Conclusions

Learning outcomes