Ultrasound-Nerve-Segmentation

Accurately identifying nerve structures in ultrasound images is a critical step in effectively inserting a patient’s pain management catheter. Hence, this is a model that can identify nerve structures in a dataset of ultrasound images of the neck. Doing so would improve catheter placement and contribute to a more pain free future.

Dataset

The data has been provided by kaggle as part of one of their competitions which requires a model for nerve segmentation. The task was to segment a collection of nerves called the Brachial Plexus (BP) in ultrasound images. A large training set of images where the nerve has been manually annotated by humans was provided.

Ultrasound Nerve Segmentation dataset - https://www.kaggle.com/c/5144/download-all

The dataset can also be downloaded using Kaggle API

kaggle competitions download -c ultrasound-nerve-segmentation

File Description

/train/ contains the training set images, named according to subject_imageNum.tif. Every image with the same subject number comes from the same person. This folder also includes binary mask images showing the BP segmentations.
/test/ contains the test set images, named according to imageNum.tif. You must predict the BP segmentation for these images and are not provided a subject number. There is no overlap between the subjects in the training and test sets.

Network Architecture

Being an image segmentation problem , wherein, just classifying the image wouldn't solve it , but segmenting within the image should help. Hence, upsampling of CNN output has to be done to produce probability mask. Hence, U-Net can be used.
U-Net has two paths - Contraction Path and Expansion Path.
Contraction path extracts context of the image and doing so, the image is down sampled.
Then, the Expansion path upsamples the image and outputs a probability mask of same size as that of input image.

Training

The images are prprocessed and resized to (128 , 128).
The model is compiled using Adam optimizer , binary crossentropy as loss and accuracy as metrics.
The model is trained for 50 epochs with Early stopping , in case the validation loss doesn't lower further.
An accuracy of around 98% is achieved by the model on the test data set.

To-do List