Leifeng network news, on December 19, 2021, the final of the ISICDM 2021 Medical Image Segmentation Challenge was successfully held, and the team of Professor Yue Xiaodong from the School of Computer Engineering and Science of Shanghai University, the team of Professor Yang Wei of the School of Biomedical Engineering of Southern Medical University, and the team of Professor Qi Shouliang of the School of Medicine and Bioinformatics Engineering of Northeastern University won the championship in a total of 2 major events and 5 small events.
The challenge was hosted by the ORGANIZING Committee of ISICDM 2021 Conference, Tianjin Medical University General Hospital, Tongji University and Northeastern University, with Professor Jiang Rongcai of Tianjin Medical University General Hospital, Professor Li Chunming of University of Electronic Science and Technology of China, Professor Yang Huihua of Beijing University of Posts and Telecommunications as the guide of the challenge, and Associate Professor Chen Yufei of Tongji University and Professor Qin Wenjun of Northeastern University as the chairman of the challenge, aiming to provide a platform for technical exchange and display for the digital medicine industry, academia and research community through the form of the challenge. The features of this challenge are as follows:
(1) The schedule lasted three weeks and the races were intertwined
Registration for the competition began at the end of November, with 43 teams from 22 institutions registering for the competition. The schedule is divided into 3 stages: warm-up, qualifying and final, each of which is uploaded for a limited time after the data is issued. There are 2 major competition items in this competition, and 5 competitions are divided: skull CT hematoma segmentation and hematoma volume calculation, lung CT anatomy segmentation - bronchi / lung lobe / pulmonary vessel / pulmonary artery and vein.
Objective indicator data are recorded at each stage of the competition process and the results of the current ranking are announced. The ranking was based on a weighted average of all objective indicators at each stage, and 12 teams reached the final.
(2) The final process is fair and transparent, and it is both professional and practical
Due to the epidemic situation, the final was held in the online conference room and broadcast live, and the final stage received widespread attention, in addition to the people who entered the conference room to watch the game, more than 4,000 people watched the live broadcast online.
The judges are professional and well-constituted: the three categories of judges are from hospitals, companies and universities, and the results of each team are comprehensively evaluated from the perspective of clinical, industry and algorithm. At the scene, professors Jiang Rongcai and Dr. Sha Zhuang from the General Hospital of Tianjin Medical University from the medical community, Dr. Xue Hongsheng of Zhongshan Hospital Affiliated to Dalian University, Dr. Zhou Di of the Fourth Affiliated Hospital of China Medical University and Dr. Xu Xinfeng of the Department of Thoracic Surgery of Jiangsu Provincial People's Hospital evaluated the results from a clinical perspective, Dr. Zhou Qinghua from Neusoft Medical guided the results from the perspective of the medical industry, and Professor Qin Wenjun from Northeastern University and Associate Professor Chen Yufei of Tongji University evaluated the innovation of the algorithm.
The evaluation angle of the whole review process covers a wide range, including algorithm accuracy, clinical availability, interface display and interactivity, method innovation, algorithm efficiency, etc., and the review angle is practical and professional.
Before the competition of project one cerebral hematoma, Professor Jiang Rongcai, deputy director of the Department of Neurosurgery and director of NICU of the General Hospital of Tianjin Medical University, deputy director of the Tianjin Institute of Neurology, and a front-line expert in severe neurological disease in China, made a profound exposition on the clinical significance of this competition.
Director Jiang is a member of the Shenwai Branch of the Chinese Medical Association and the deputy leader of the Brain Trauma Group, the vice chairman of the Neurological Intensive Care Committee of the Shenwai Branch of the Chinese Medical Doctor Association and the vice chairman of the Cranial Trauma Committee of the National Trauma Medical Center, as well as the Chinese Journal of Neurosurgery. Editorial board member of the Chinese Journal of Anatomy and Clinical, Journal of Clinical Neurosurgery and J Clin Lab Anal. He is one of the two inventors of atorvastatin calcium treatment of chronic subdural hematoma and optimized the treatment plan, and is the main contributor of the project to win the Tianjin Science and Technology Progress Special Prize.
Known for its innovative concept for the treatment of severe brain injuries and their complications, it focuses on the regulation of intracranial lymphatic drainage and the treatment of brain trauma, and has treated more than 1,000 patients with severe neurological diseases every year. He is the head of the innovation team of Tianjin 131 First-level Talents, the 2020 Tianjin Key Areas Promotion Plan, and the "Famous Doctor of the Country. Excellent demeanor" and "top ten original research leaders" to change medical practice, founded the Tianjin Medical Association Neurological Intensive Care Branch.
Director Jiang pointed out that cerebral hematoma measurement has important clinical significance in surgical treatment, and one of the surgical indications for cerebral hemorrhage patients in clinical practice is the amount of cerebral hematoma, such as patients with suprate hematomas greater than 30mL or subsurfaceal hematomas greater than 10mL need surgery, but clinically it is impossible to accurately calculate the volume of cerebral hematomas.
Skull CT examination can show the bleeding lesions very well, and accurately estimating the amount of hematoma in CT images is an urgent problem to be solved in various studies, and the hematoma results manually segmented by the physician are currently used as the "gold standard" for calculating the amount of hematoma, but this method is time-consuming and laborious; clinically, the Tada formula is often used as an estimation method for hematoma volume, which is more accurate for hematomas with relatively regular bleeding shapes, but when the bleeding shape is relatively irregular, the results are difficult to say accurate. In recent years, with the advancement of artificial intelligence (AI) technology, according to the CT image characteristics of cerebral hematoma, it has become possible to automatically segment the hematoma area in the CT image and calculate the volume of the hematoma, and this competition intends to develop accurate and convenient cerebral hematoma segmentation and hematoma volume calculation methods to provide clinical technical support for the accurate diagnosis of cerebral hemorrhage.
The competition process emphasizes actual combat: in terms of practicality, the medical image training data and test data of the two competition items are from the real clinical data provided by the General Hospital of Tianjin Medical University and Northeastern University respectively.
The competition does not allow repeated submission of split results, avoiding the drawbacks of the participants brushing scores through repeated submission of results in the traditional split challenge; the time of each game in the preliminary and final sessions is getting shorter and shorter, and it is getting closer and closer to the real clinical environment; in the final link, the challenge organizers issue clinical data on the spot, 1 person defends and shares PPT, 1 person tests the data released on the spot, and the judges and the audience can see the entire process of software operation through the contestant screen, which is real and credible.
Among them, the warm-up race and qualifying match take a total of two hours from the issuance of test data to the submission. The final stage of the on-site distribution of test data and demonstration need to be completed within 10 minutes. The final stage was more intense than the warm-up and qualifying rounds, and to some extent ruled out the possibility of manual revision.
Rigorous and transparent follow-up of the schedule: the final result of the competition comprehensively takes into account the results of all objective indicators weighted averages of each stage of the team, as well as the subjective evaluation results of the judges of the live final. The conference team announced the calculation and quantitative indicators of each project in advance, and disclosed the ranking of the teams in each stage and other information, which is convenient for the participating teams to verify the accuracy of the ranking.
In addition, Wu Qian, Fu Wei, Sun Shichen, Huang Qiguang from Tongji University and Wang Longguang, Li Xiaoshuo, Huang Peifang and Zhou Luyu from Northeastern University, 8 challenge assistants, have certain medical expertise and algorithm professional ability, follow up and supervise the whole course of the competition, and also provide a guarantee for the smooth progress of the challenge.
(3) The participating teams have strong strength and the highlights of the model are frequent
In this competition, the participating teams demonstrated strong algorithmic strength and in-depth thinking on clinical problems. Aiming at the inaccurate data annotation and segmentation model selection problem in the CT image segmentation of cerebral hematoma, Yue Xiaodong's team at the School of Computer Engineering and Science of Shanghai University expanded and improved the existing CNN and Transformer segmentation models, proposed a segmentation method based on global context attention, and adopted a multi-model integration strategy to measure and segment the hematoma region, ensuring that the segmentation prediction results have good robustness and generalization performance, and to a certain extent, making up for the lack of labeling by doctors. In the test of real-time data in the field, it has been well received by various experts.
The Yang Wei team from the School of Biomedical Engineering of Southern Medical University has achieved good results in the competition of both projects. For the task of segmentation of lung CT anatomy, they proposed to use multi-scale and lightweight UNet to solve the problems of fine vascular segmentation and arteriovenous segmentation requiring larger receptor fields, respectively, and for the segmentation of head CT hematoma and the calculation of hematoma volume, the integration method of nnUNet and nnFormer was used to improve the accuracy of hematoma segmentation and hematoma volume estimation, which was outstanding in various indicators.
Qi Shouliang's team from northeastern university's School of Medicine and Bioinformatics Engineering adheres to the idea of "one architecture fits all segmentation" and realizes the anatomical segmentation of lung lobes, airways, blood vessels, arteries and veins of the lungs based on the 3D-UNet model. In the decoder part of the model, a deep supervision method is used to integrate the output results of different resolutions to generate the prediction results of the model, which has achieved good results in the segmentation task of various anatomical structures of the lungs.
The 2022 ISICDM Segmentation Challenge will continue to maintain the strengths and features of this year's Challenge, inviting more algorithm experts and doctors to participate in the data collation and release of the Challenge, formulating the schedule and rules, and serving as the challenge judges. It is believed that with the addition of more experts and scholars, next year's ISICDM Challenge will be more exciting, and is expected to promote the research and development and landing of related technologies.
Full list of winners: