A Combined Neural and Temporal Approach for Tracking Anatomical Features in Liver.

Mélanie Bernhardt - ETH Zurich - May 2019

Abstract

Ultrasound motion tracking is required for various medical applications. In this report, we describe a combined approach for tracking anatomical landmarks in liver during respiration, based on the CLUST Challenge. The proposed method combines a local Siamese-CNN and a Ridge Regression temporal model for feature localization at each frame. The method was developed and fine-tuned via 5-fold across-sequence cross-validation and then evaluated on the CLUST Challenge Test set.

About this repository

This repository contains the code related to the project. The report describing the methods and the results can also be found under report.pdf.

Setup

In order to run any code of this repository 3 environment variables have to be set:

• EXP_PATH the path to the directory saving the checkpoints
• DATA_PATH the path to the training data
• TEST_PATH the path to the testing data.

Main files

• To run cross_validation evaluation use global_tracking.py. Set the parameters you wish to use and the name of the experiment in the parameters dictionary at the end of the file and run.
• To train, predict, visualize the predictions simply use the jupyter notebook cells in train_predict_visualize.