Generative Models Based Medical Image Classification And Adversarial Attacks

Medical imaging plays a crucial role in the diagnosis and treatment of patients.It helps medical experts make crucial decisions about patients,such as whether cancer patients should be treated with chemotherapy or surgery.Even if the patient survives in crucial surgery,such decisions greatly impact the quality of the patient’s life for the rest of his life.So,medical experts make such decisions after several medical tests with complete responsibility,and it adds the burden on already overloaded healthcare systems.Extracting the information efficiently and quickly from the medical images plays an important role in making clinical decisions.Due to the heavy workload on healthcare systems,medical experts are prone to make wrong decisions about the patient’s diagnosis.Therefore,adopting computer-aided diagnostic(CAD)tools would improve the performance of medical staff and help reduce the burden on the healthcare system.Machine learning methods have been adopted to improve the performance of medical image applications.However,applying traditional machine learning methods in medical image classification depends on feature extraction and selection methods that are sensitive to different colors,shapes,and sizes.Deep learning models have become a successful and promising alternative to overcome the drawbacks of traditional machine learning methods that use handcrafted features.The deep learning computing paradigm has been deemed the gold standard and most widely used computational approach in the machine learning community.It is able to match or even beat those diagnosis results provided by a human.Even though deep learning has greatly contributed to the advancement of the field in solving different tasks,we found several drawbacks and pitfalls that need to be addressed,especially when dealing with medical imaging tasks.The exceptional performance of deep learning usually requires a huge amount of labeled data which is a crucial part of medicine.Medical image datasets are rare as they require more time and resources to collect and label than any other data collection task.Patient privacy is also the main hurdle to collecting medical data.So finding ways to handle the limited availability of medical datasets in various imaging domains of medicine is of great significance and need of the time.Moreover,recent research has revealed that deep learning models,even the modern ones,are susceptible to adversarial examples despite their superior performance.Adversarial examples are also called adversarial attacks in which perturbed input instances,even the slightest ones,can fool deep learning models to make incorrect predictions with a significant degree of confidence.It raises major safety concerns about using artificial intelligence-based systems in critical real-life domains such as medicine.In this research,we deal with both the mentioned problems by proposing a generative modelbased solution for limited medical datasets and a novel adversarial example detection method.The main contributions of the thesis are as follows:1-Recently,generative adversarial networks(GANs)received the attention of researchers around the world.GANs are famous for generating realistic-looking images from natural and medical imaging.These generated images are used to balance the imbalance dataset or as a data augmentation in various studies where datasets are small.GANs generator takes the fixed-length random vector consisting of noise as an input and generates a sample image while the discriminator distinguishes the real and generated sample images.Researchers mostly use the Gaussian random noise as input and try to improve the GAN’s results by using other aspects of the GANs.The area of noise that GAN use for the generation of images is unexplored.We leveraged this and manipulated the noise model of GAN.We use the autoencoders network,train it on the training dataset to let it learn the noise distribution of the domain,and let the GAN sample the noise vector from this informative noise instead of using the random noise.In this way,one major problem of generative adversarial networks,mode collapse solved.Experimental results show that the GANs convergence time is also reduced using the proposed method.Moreover,the proposed method is general;it can be used in any reallife problem where the image dataset is small or difficult to collect.2-We improve the skin cancer classification by using generative adversarial networks and heavy-tailed student t-distribution.In this part of the research,we use the combination of image processing,generative adversarial network and autoencoders.We trained the autoencoder network on the training dataset and swapped the encoder-decoder network with the decoder-encoder network,and used the GAN to input the noise vector from the output of the decoder-encoder network.The images generated by this GAN were used(along with the training set)to train the discriminator of the primary generative model.Moreover,the generator of the primary network uses the student t-distribution instead of Gaussian random noise.The heavy-tailed t-distribution helps the generator to produce diverse images.And finally,we use the high pass filter on the generated images of the primary GAN.Comparative results with several generative models show that the proposed method has significance in producing high-quality and diverse images.We tested the proposed methodology on the skin cancer classification dataset.3-Adversarial attacks on medical images.Deep learning models are now the foundation of many image analysis and image processing-based studies.Although deep learning performs exceptionally well on vision tasks,current research indicates that it may be subject to adversarial attacks.Most of the research related to adversarial attacks is based on natural images.We discussed that medical images are quite a different domain than natural images.It is also important to explore the various aspects of adversarial attacks on medical images.In this study,we implemented three adversarial attacks on two medical imaging datasets and observed the effect of generative and public datasets on attack transferability.i-e How is it likely to transfer the adversarial attack when the surrogate model is trained on the public dataset and the target model is trained on the generative(private)dataset? Moreover,we observe the effect of various levels of perturbations on medical images and observe the reduction in the performance of the trained models.Our experimental results show that medical images are more sensitive to adversarial perturbation than natural images.4-Negative image representation-based adversarial example detection in medical imaging.In literature,some methods are proposed to reduce the effect of adversarial attacks.Adversarial training of the model is one of the most commonly used ones.But attacks keep updating the adversarial attacks,or they use the combination of the adversarial attacks to fool the target network.In such situations,detection mechanisms are more valuable than the defense mechanism.So,we propose a novel technique for the detection of adversarial attacks in this study.We leverage the negative image representation(NRI)inspired by the biological immune system to serve the purpose.We include negative samples of input instances in our training process and observe exceptional adversarial example detection results.We tested the proposed method on three medical datasets against five different adversarial attacks.Experimental results show that the proposed method has significantly better results than the state-of-the-art methods.