Privacy preservation in medical image analysis

Abstract

In recent years, the development of artificial intelligence (AI) algorithms has been the subject of tremendous progress brought namely by the rise of deep neural networks. Research works in AI have been translated into a broad range of applications, particularly, in the field of computer vision and natural language processing. In contrast, medical applications of AI have been slower to appear until now, largely due to privacy constraints on medical data. The primary objective of this thesis is to develop novel deep learning methods for client-server medical image analysis, which also protect patient privacy. Toward this goal, three methods are proposed, each one associated with a distinct thesis chapter. These methods follow the same high-level strategy where a client (for instance, a hospital) encodes an image so that the sensitive information is obfuscated, before sending it to a server for analysis. The server processes the encoded image with a neural network designed to handle this data, and then sends the result back to the client where it is finally decoded. An important constraint to satisfy in this setting is never sending information that can be used to identify patients. The first method, based on adversarial learning, is composed of three networks trained end-to-end: an encoder network which removes identity-specific features from the input image, a discriminator network that attempts to identify the corresponding subject from the encoded image, and a segmentation network which tries to extract regions of interest in the same image. By training the encoder to simultaneously fool the discriminator and maximize segmentation performance, it can learn to remove private features while keeping those essential for the segmentation task. The method’s ability to provide a high-quality segmentation, while also obfuscating patient identity, is demonstrated on the segmentation of brain MRI from the large-scale Parkinson Progression Marker Initiative (PPMI) dataset. A limitation of the first approach is that it does not encode the segmentation maps produced by the network, which may also be used to identify the patient. The second method presented in the thesis, which aims to alleviate this problem, protects patient identity by applying a pseudorandom non-linear deformation to the input image. This results into a proxy image which is sent to the server for processing. The segmentation network then produces a deformed segmentation map in which the patient’s identity is protected. This map is sent back to the client where it is reverted back to a canonical form. The overall system has three components: a flow-field generator which produces a pseudo-random deformation function, a Siamese discriminator that tries to recover the patient identity from the processed image, and a segmentation network that analyzes the content of the proxy images. As in the first approach, the system is trained end-to-end in an adversarial manner. By fooling the discriminator, the flow-field generator learns to produce a reversible non-linear deformation which allows to remove information related to patient identity from both the input image and resulting segmentation map. The proposed method is once again validated on the task of MRI brain segmentation using images from two different datasets. Results show this method to offer a segmentation accuracy similar to a system trained on non-encoded images, while also reducing considerably the ability to recover subject identity. The first two approaches use an adversarial learning strategy based on Siamese discriminators to encode the images to protect. However, training such discriminators requires to have several images for the same patient, which is not always possible in practice. Moreover, the adversarial training of the model is often unstable and the trained model can be sensitive to changes in the distribution of images. The last method explores a different strategy to preserve privacy. In this method, the client protects the to-be-segmented patient image by mixing it to a reference image, making it unworkable and unrecognizable for an unauthorized person. This proxy image is sent to a server for processing. The server then returns the mixture of segmentation maps to the client, which can revert it to a correct target segmentation. The system has two components: a segmentation network on the server side which processes the image mixture, and a segmentation “unmixing” network which recovers the correct segmentation map from the segmentation mixture. Once more, the whole system is trained end-to-end. The results of experiments on different datasets show that this method achieves a high or comparable segmentation accuracy with respect to previous approaches, while also being simpler to implement and requiring less computations than these approaches.

Date	21 Dec 2023
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Christian Desrosiers (Supervisor), José Dolz (Co-supervisor) & Pierre Marc Jodoin (Co-supervisor)