Medical image segmentation with limited data

Abstract

[en] Ischemic Heart Disease (IHD) is one of the leading causes of mortality in Spain; early diagnosis is key. Intravenous ultrasound imaging (IVUS) can help identify symptoms of IHD, at the cost of segmenting a large volume of frames by medical professionals. While promising, automated image segmentation using Convolutional Neural Networks (CNN) suffer from sample scarcity: a large amount of parameters is often used, and medical imaging datasets are typically small and costly to acquire and label. In this report we study and compare state of the art methods used to deal with sample scarcity. In particular we introduce data augmentation methodologies, specialized training losses and transfer learning methods, and compare their performance on IVUS segmentation of the media and lumen or the artery. Additionally we introduce a promising paradigm, few-shot segmentation, and provide an initial implementation using PFENet. This implementation can avoid significant overfitting, even when trained with a single example, outperforming traditional CNNs on the same segmentation problem.