Deep learning models based on automatic labeling with application in echocardiography
Manuela Daniela Danu, Costin Florian Ciusdel, Lucian Itu · 2020
Our goal is to apply deep learning (DL) based techniques in the context of medical imaging in order to perform automatic data labeling. Generally, training a neural network requires large amounts of (annotated) data, which is often very difficult to obtain, especially in the medical field. To alleviate this problem, we propose the use of pre-trained models. In this paper, we present two pre-training methods, formulated as self-supervised classification problems. The first one is a binary classification problem, which aims to label echocardiographies according to how they are represented: flipped or non-flipped. For this first task, we achieved 95.83% accuracy. In the second approach we aim to sort into chronological order a tuple of three randomly shuffled frames from a cardiac ultrasound DICOM file. Even though DICOM videos have no semantic labels, we use the temporal dimension as a supervisory signal. Since for each tuple of three frames there are 6 possibilities of arranging them, the sequence sorting task is formulated as a multi-class classification task with 6 classes. For this task we obtained 92.38% operating accuracy. In order to improve the learning performance for this second task, we employ transfer learning methods by re-using learned layers from the first classification task. As a result of the knowledge transfer process, the accuracy increased to 95.43%. This way, we showed that transfer learning and self-supervision hold the potential to yield significant improvements in the learning process of the target task.