PRACTICAL FEASIBILITY OF DEEP-LEARNING BASED MEDICAL IMAGE SEGMENTATION FOR THE AORTA

Abstract

Medical imaging segmentation can help health workers analyze and examine abnormal changes in an organ or in tissue due to injury or disease, without invading the patient's body. The application of deep learning makes this process more efficient. There are many studies have been presented in this field, especially in the segmentation of liver, heart, kidney and lung with very successful results. However, there are still many challenges of using deep learning on other organs' segmentation, such as the aorta. This report presents the feasibility of using deep learning on aortic segmentation. Firstly, we explored different methods for generating accurate ground-truth segmentation as training data, concluded the average time cost for each scan is around 3 hours. Secondly, we compared to the segmentation result of kidney using a large training set, analyzed the practical constraints which prevent us for getting an equivalently good result, including the training time (on the order of hundreds of hours), and the cost for accessing the computing hardware is also on the order of thousands of US dollars. These practical constraints are usually not disclosed by those successful studies, but are crucial for those researchers who would like to perform the deep learning approach to the medical image segmentation. This report summarizes these challenges and provides lessons learned for future practitioners.