In this project, we propose an algorithm for automatic segmentation of nerves in ultrasound images of the neck. Deep convolutional neural networks, when trained end-to-end, pixels-to-pixel, have demonstrated incomparable performance for various tasks such as image classification and object detection. Image segmentation is a low-level vision task which involves pixel level classification. Recent techniques which used fully convolutional neural networks for semantic segmentation have limited the scope for Image Segmentation. In this project, we demonstrate the use of the U-Net convolutional architecture for ultrasound nerve segmentation.
The dataset contains 5635 training images with their correspomding masks and 5508 test images who masks are supposed be predicted. The data is sourced from the Kaggle, and similarly sized test dataset. The test dataset would be used to detect the presence of Brachial Plexus nerve in the images in an automated manner.
The dataset consists of 5635 training images from 47 patients and their corresponding masks. The testing data consists of 5508 images which are to be segmented. It is important to separate the training images and their masks in different arrays. The images were resized from 480520 to a size of 6480. It is important to keep all the patient data together with a view to avoid overfitting. Image data from 42 patients was used to trained the network. Images of 5 patients were kept as validation. The test data is the unseen data and validated will predict the mask for the test images. The preprocessing also consisted of normalization of data to avoid imbalance in weight updates. Image pixel values were between [0,1] indicating grayscale and mask pixels were converted to 0 or 1 indicated segmented image. The 3-fold cross validation used initially was not used later with a view to train on a lager dataset and restrictions on computing power.
Neural networks are a powerful tool for data classification. Convolutional Neural Networks can be used for Image classification. The same technique can be extended to Image Segmentation which is nothing but pixel-wise classification. The U-Net architecture is a very useful technique for biomedical image segmentation due to the complexity of biomedical images. However, due to quality and amount of training data and constraints on computing power, the achieved value of dice coefficient is low. However, with better computing power and optimization of the neural network the Images can be better segmented and construction of a model with greater accuracy will be possible. The only downside of convolutional neural networks is that it requires huge training data and large training time.
- Disha Doshi
- Sadaf Ghaffari
- Ghananeel Rotithor
- Piyush Agade