prashantkh19 / Ultrasound-Nerve-Segmentation

Identify nerve structures in ultrasound images of the neck

Ultrasound-Nerve-Segmentation

Identify nerve structures in ultrasound images of the neck

The task is to segment a collection of nerves called the Brachial Plexus (BP) in ultrasound images. A large training set of images is used where the nerve has been manually annotated by humans. Annotators were trained by experts and instructed to annotate images where they felt confident about the existence of the BP landmark.

Dataset

The dataset has been acquired from the Kaggel Competition -Ultrasound Nerve Segmentation.

References

The research paper, U-Net: Convolutional Networks for Biomedical Image Segmentation by Olaf Ronneberger, Philipp Fischer, Thomas Brox is used as a reference.

Code

The code is written using fastai modules based on pytorch.

Kaggle Kernel is used to code the whole project. The notebook can be viewed here.

Environment Details:

To be updated

Configuration

ResNet-34 encoder + U-Net

Original U-Net architecture:

This is how the original U-Net looks like.

I have used ResNet-34 encoder (pre-trained on Image-Net dataset) on U-Net base.

Reason for the modification:

Taking pre-trained resnet encoder gives the whole training a head-start in understanding the basic image properties.

Results:

• Achieved top 10% of the competition's leaderboard on Validation Set with a dice score of around 0.7.

Loss Plot

The above results are obtained with about half-an-hour training of the chosen architecture.