Robust Deep Learning-based Computer-aided Diagnosis Against Erroneous Data

Abstract

Recently, computer-aided diagnosis (CAD) has been extensively paid attention by researchers to support physicians. CAD is defined as a system which assists doctors to interpret medical data such as computed tomography (CT) scans, electroencephalography (EEG), X-ray, magnetic resonance imaging, electromyography, electrocardiography etc.

Since the success of the image recognition using deep learning model, CADs based on the deep learning have been tremendously emerged for various tasks. Moreover, researchers confirm that CADs using deep learning model achieve comparable performance to human experts. Therefore, the robustness and explainability are more considerable than the performance to avoid system failure. There are three reasons which can cause system failure of the CAD based on deep learning as follows. (1) Erroneous data which is caused accidentally can occur a system failure of CAD based on deep learning. (2) Erroneous data such as adversarial samples by intended intruders can cause a system failure of CAD based on deep learning. (3) Diagnosis by CAD is not reliable because CAD based on deep learning is considered a black box which is not explainable while the erroneous data are occurred. Therefore, the objective of this work was to develop a robust CAD based on deep learning which is against the erroneous data.

In this work, the robustness of the CAD is developed in the scenarios as follows. (1) While missing values are very prevalent in medical domain and it is originated by various reasons including human error or device failures which are not intended, it causes the critical problem in medical application such as not only incorrect diagnosis but also system failure. Therefore, to deal with the erroneous data, the first problem of this work is to develop a system which supports a deep learning-based CAD to be robust against missing values. (2) In addition, it was revealed that deep learning-based models are vulnerable from adversarial attacks. However, because existing adversarial attacks are based on additional noise which is called perturbation, it is easily captured by the physician. Consequently, existing adversarial attack is not appropriate to measure robustness of deep learning-based CAD. Thus, to handle the erroneous data, the second problem of this work is to develop an adversarial attack which is imperceptible. (3) Finally, although deep learning-based CAD performs comparable performance to human experts, it is difficult to replace them yet because it is a black box. Existing interpretation methods using Local Interpretable Model-agnostic Explanations (LIME) depends on clustering method of input instance which are randomly clustering without considering criteria of diagnosis by physicians. Hence, to enhance robust of a CAD against the erroneous data, the third problem of this work is to develop a method to provide an explanation whether a CAD based on deep learning has same criteria of physicians.

To tackle the problems of the motivations, this work solves the problems as follows. The first problem is handled with generative adversarial network acquiring the contextual information while the existing works ignore it. Moreover, the second problem is tackled by the adversarial attack distorting of features based on encoder-decoder model. In addition, the third problem is approached by novel instance split method using short-time Fourier transform and inverse it and acquiring LIME. In experiments, the suggestions are evaluated using real world dataset which are computed tomography and electroencephalography. In experiments, the suggestions outperform the state-of-the-art methods. Additionally, the comprehensive comparisons are provided from the experimental results.

This work enhances the robustness of the computer-aided diagnosis based on deep learning against erroneous data of the missing imputation system, the adversarial samples using the explainable CAD. In addition, it is believed that this work promotes the deep learning-based CAD in practice in the future.