Shuchao Pang1, Anan Du2, Mehmet A Orgun3,4, Zhenmei Yu5, Yunyun Wang6, Yan Wang1, Guanfeng Liu1. 1. Department of Computing, Macquarie University, Sydney, NSW, 2109, Australia. 2. School of Electrical and Data Engineering, University of Technology Sydney, Ultimo, NSW, 2007, Australia. 3. Department of Computing, Macquarie University, Sydney, NSW, 2109, Australia. mehmet.orgun@mq.edu.au. 4. Faculty of Information Technology, Macau University of Science and Technology, Avenida Wai Long, Taipa, 999078, Macau, China. mehmet.orgun@mq.edu.au. 5. School of Data and Computer Science, Shandong Women's University, Jinan, 250014, China. 6. Department of Anesthesiology, China-Japan Union Hospital of Jilin University, Changchun, 130012, China.
Abstract
PURPOSE: Unlike the normal organ segmentation task, automatic tumor segmentation is a more challenging task because of the existence of similar visual characteristics between tumors and their surroundings, especially on computed tomography (CT) images with severe low contrast resolution, as well as the diversity and individual characteristics of data acquisition procedures and devices. Consequently, most of the recently proposed methods have become increasingly difficult to be applied on a different tumor dataset with good results, and moreover, some tumor segmentors usually fail to generalize beyond those datasets and modalities used in their original evaluation experiments. METHODS: In order to alleviate some of the problems with the recently proposed methods, we propose a novel unified and end-to-end adversarial learning framework for automatic segmentation of any kinds of tumors from CT scans, called CTumorGAN, consisting of a Generator network and a Discriminator network. Specifically, the Generator attempts to generate segmentation results that are close to their corresponding golden standards, while the Discriminator aims to distinguish between generated samples and real tumor ground truths. More importantly, we deliberately design different modules to take into account the well-known obstacles, e.g., severe class imbalance, small tumor localization, and the label noise problem with poor expert annotation quality, and then use these modules to guide the CTumorGAN training process by utilizing multi-level supervision more effectively. RESULTS: We conduct a comprehensive evaluation on diverse loss functions for tumor segmentation and find that mean square error is more suitable for the CT tumor segmentation task. Furthermore, extensive experiments with multiple evaluation criteria on three well-established datasets, including lung tumor, kidney tumor, and liver tumor databases, also demonstrate that our CTumorGAN achieves stable and competitive performance compared with the state-of-the-art approaches for CT tumor segmentation. CONCLUSION: In order to overcome those key challenges arising from CT datasets and solve some of the main problems existing in the current deep learning-based methods, we propose a novel unified CTumorGAN framework, which can be effectively generalized to address any kinds of tumor datasets with superior performance.
PURPOSE: Unlike the normal organ segmentation task, automatic tumor segmentation is a more challenging task because of the existence of similar visual characteristics between tumors and their surroundings, especially on computed tomography (CT) images with severe low contrast resolution, as well as the diversity and individual characteristics of data acquisition procedures and devices. Consequently, most of the recently proposed methods have become increasingly difficult to be applied on a different tumor dataset with good results, and moreover, some tumor segmentors usually fail to generalize beyond those datasets and modalities used in their original evaluation experiments. METHODS: In order to alleviate some of the problems with the recently proposed methods, we propose a novel unified and end-to-end adversarial learning framework for automatic segmentation of any kinds of tumors from CT scans, called CTumorGAN, consisting of a Generator network and a Discriminator network. Specifically, the Generator attempts to generate segmentation results that are close to their corresponding golden standards, while the Discriminator aims to distinguish between generated samples and real tumor ground truths. More importantly, we deliberately design different modules to take into account the well-known obstacles, e.g., severe class imbalance, small tumor localization, and the label noise problem with poor expert annotation quality, and then use these modules to guide the CTumorGAN training process by utilizing multi-level supervision more effectively. RESULTS: We conduct a comprehensive evaluation on diverse loss functions for tumor segmentation and find that mean square error is more suitable for the CT tumor segmentation task. Furthermore, extensive experiments with multiple evaluation criteria on three well-established datasets, including lung tumor, kidney tumor, and liver tumor databases, also demonstrate that our CTumorGAN achieves stable and competitive performance compared with the state-of-the-art approaches for CT tumor segmentation. CONCLUSION: In order to overcome those key challenges arising from CT datasets and solve some of the main problems existing in the current deep learning-based methods, we propose a novel unified CTumorGAN framework, which can be effectively generalized to address any kinds of tumor datasets with superior performance.