Semi-automatic liver segmentation for ground truth database development for Deep Learning Network
Pravda Jith Ray Prasad, Informatics
Introduction: Medical image segmentation has gained greater attention over past few years, especially in image-guided surgery. Accurate and fast liver segmentation tools are important for liver resection planning and navigation. Currently we had developed an interactive semi-automatic method and propose to develop an automatic segmentation method using Deep Learning, where our goal is to improve the speed and accuracy. Our research objective is to analyze whether Deep Learning can improve liver and lesion segmentation outcomes as compared to the existing methods .
Methods and Materials: The implementation of the semi-automatic method has been done in 3DSlicer which is an open source software widely used by clinical practitioners and biomedical engineers. This segmentation work-flow starts with the user specifying the region of interest for information gathering followed by filtering. Finally, user can make corrections for refining the segmentation results.
Results and Discussion: The interactive approach has been applied to 10 different CT datasets and percent volume error is measured as 1.02% (obtained by comparing the volume difference after manual segmentation done with the help of an expert). This method will be evaluated using larger datasets so as to validate the accuracy. The results obtained will be later used as ground truth database for training the Deep Learning network. In summary, we developed a semi-automatic approach with good segmentation results in CT images, which will be further subjected for more validation. Also we propose to develop an automatic segmentation method on Deep Learning networks for the better speed and accuracy compared to existing methods.
References:  Litjens, G. et al. A survey on deep learning in medical image analysis. arXiv preprint, arXiv:1702.05747 (2017).