DeepSynthBody: the beginning of the end for data deficiency in medicine

Abstract

Hammer, for their support, motivation, and always behind me. Without them, this would be only a dream. After joining the HOST department as a Ph.D. candidate in 2018, I started experiencing a completely new environment with new people from different countries and cultures.Michael Riegler became my principal supervisor. I did not know anything about him despite his academic background. However, after few weeks, I realized that he is more than my principal supervisor for my life. Within few months, he became a game-changer in my life. I do not have words to express his qualities and how his advice, encouragement, kindness, and motivations are important to my life. Therefore, I would like to give my most enormous thanks to my primary supervisor, Michael, who is always behind me to support my academic life and get advice for my personal life.Pål Halvorsen is the department head and one of my co-supervisors, and he is a very kind person supporting us always silently. His advice and encouragements make me better and better every day in my academic life. Then, I would like to thank Pål for his kindness showed me through my Ph.D. journey and the wordless help given to me. Hugo Hammer is my second co-supervisor who supports me in handling advanced statistical problems in my Ph.D. research. Then, I would like to thank Hugo for his friendly help, as always.The research has benefited from the Experimental Infrastructure for Exploration of Exascale Computing (eX3), which the Research Council of Norway financially supports under contract 270053. Tore was the person who helped me to use these infrastructures.So, I would like to give my special thanks to him for his outstanding support.I also would like to thank the closest colleague, Steven Hicks, a good friend met within the department and in my life. Moreover, I thank all the other colleagues, Debesh, Hanna, Pia, Hakon, and Daniel, who work closely in my Ph.D. periods. There are more people v I want to thank, but writing them all here is not possible. However, they all are equally important for me, and therefore I would like to thank them all equally. Master students, other Ph.D. students from different departments and other counties, all the co-authors, and persons who are from the administration are a few of them.Finally, I would like to thank my wife, Shalike, who sacrificed the freedom of her life for finishing my Ph.D. successfully. Not only that, she always provided me food and other necessary stuff without asking anything from me. Last but not least, I would like to give my thanks to my parents, who sacrificed their whole life to provide us a good life while facing a challenging lifestyle in my country,

Application of Example-Based Explainable Artificial Intelligence (XAI) for Analysis and Interpretation of Medical Imaging: A Systematic Review

“…In parallel, the study by Vajira Thambawita et al [13] represents a milestone in using GANs to develop XAI within the medical domain. DeepSynthBody not only tackles the problem of limited data in the healthcare sector by generating synthetic data using advanced GANs but also integrates a crucial explanatory dimension.…”

“…GAN-based techniques such as cGANs and DeepSynthBody provide a global perspective by generating synthetic images. cGANs excel at generating hypotheses and visual insights into image characteristics, while DeepSynthBody distinguishes itself by integrating explainability into all stages of the generation process, reinforcing confidence in model results [12], [13]. In contrast, approaches such as SCOPe and TraCE, which do not rely on GANs, focus on creating high-quality counterfactuals, demonstrating versatility and skill in identifying shortcuts in complex models [14], [15].…”

“…At the same time, SCOPe stands out for generating counterfactuals close to the data manifold [8], [9], [14]. DeepSynthBody raises the bar by integrating explainability into the synthetic data generation process, increasing confidence in machine learning solutions in the medical domain [13].…”

Artificial Intelligence for Colonoscopy: Past, Present, and Future

“…The most common SDA approach for colonoscopy applies user-specified image transformation methods such as rotation, zooming in/out, and cropping and translation [165]. SinGAN-Seg [166] is a generative adversarial network-based method recently proposed to generate synthetic images for polyp segmentation. • Active Learning (AL): Given a small initial training dataset, AL methods minimize manual labeling efforts by using a query strategy to select necessary sample images (typically from an unlabeled dataset) for the domain experts to classify.…”

PolypConnect: Image inpainting for generating realistic gastrointestinal tract images with polyps

“…However, producing realistic synthetic polyps with the corresponding ground truth, which can be used to train other machine learning models, is challenging. Random synthetic GI-tract images can be generated from the pre-trained generative adversarial network (GAN) models studied in [8], [9]. However, generating synthetic polyps and corresponding ground truth is not possible with these model.…”