De-identification of medical images using object-detection models, generative adversarial networks and perceptual loss
Abstract
Medical images play an essential role in the process of diagnostics and detection of a variety of diseases. Whether it being anatomical features or molecular cells, medical imaging help visualize and gain insight into the human body. These images are a crucial aid in the process of diagnosing patients. While these images are informative, they can also be quite difficult to interpret, necessitating highly trained medical professionals to read the images. The amount of medical images produced is enormous compared to the amount of professionals whose task it is to interpret them. The diagnosis can also vary based on the medical professional who inspects the image. The recent rise of a new generation of Computer Aided Detection (CAD) systems based on machine learning has become more and more important to battle this problem. These systems aids the medical professional in the diag- nostic process. This can lead to a more consistent and accurate interpretations of medical images by removing some human bias. In addition, such systems can be used to decrease the workload by either filtering out images deemed as belonging to healthy subjects, to be otherwise not of interest, or marking images as indicating a risk. When creating CAD systems utilizing machine learning you are very de- pendent on data. Since the systems will typically be placed in very delicate, high-risk situations, the quality of the data is always a priority. A common problem in medical imaging research is not getting sufficient data. Not that there is a shortage of images, but to be used in research, they typically have to be de-identified or anonymized. This process has to be verified manually and is therefore time-consuming. With the impressive advancement of machine learning in recent decades, it seems natural to attempt de-identification using machine learning, especially because several powerful models are being applied to similar tasks in other fields. One key reason for the success of machine learn- ing is its ability to detect and generate patterns. Currently, there are several applications that perform de-identification by placing black-boxes on top of in- formation detected as being sensitive [1, 2]. However, the black boxes can end up hiding also other parts of the image, but ideally all non-sensitive features in the image should be preserved. In this thesis we investigate the effect of using image-to-image deep learning to automate 2D medical image de-identification by detecting the sensitive information, and removing it without the use of black boxes. Our results indicate that de-identification models based on machine learning can result in viable and powerful solutions. The deep learning models manage to accurately detect and remove text, without large negative impact on the original image.