Research Findings of NEU’s Key Laboratory of Medical Image Intelligent Computing of MOE Published in Top International Academic Conferences and Journals on Medical Image Computing
Recently, NEU’s Key Laboratory of Medical Image Intelligent Computing of MOE has made new progress in the analysis of cardiac structure and functions, and the diagnosis and treatment prediction of mental diseases, and has published a series of six papers in top international academic conferences and journals on medical image computing. Among them, two papers were accepted by Medical Image Analysis (MedIA), three by MICCAI2023 and one by IJCAI2023. The results of these papers are of great significance to accelerating the clinical application of medical image AI and improving the efficiency of medical diagnosis. MedIA is the top journal in the field of medical image analysis. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) is a comprehensive top-level academic conference organized by Medical Image Computing and Computer Assisted Intervention Society, which is currently recognized as the most influential international conference in the field of medical image analysis. International Joint Conference on Artificial Intelligence (IJCAI)is one of the top international academic conferences in the field of AI (CCF-A), with an acceptance rate of about 15%.
In the aspect of intelligent analysis of cardiac structure and functions, two works from NEU have been accepted by MedIA, namely, Learning What and Where to Segment: A New Perspective on Medical Image Few-Shot Segmentation and Automatic Identification of Septal Flash Phenomenon in Patients with Complete Left Bundle Branch Block. Doctoral students Qu Mingjun and Feng Yong are the first authors of the two papers respectively, and Professor Yang Jinzhu is the corresponding author. This study proposed a small sample image segmentation method based on medical prior knowledge. In this study, unsupervised registration and multi-scale prototype were used to excavate the spatial and semantic features of limited supporting samples and compress the hypothesis space. It has effectively alleviated the bias and class imbalance in small sample segmentation of medical images, and improved the ability of small sample segmentation model to quickly adapt to new targets. Additionally, a cascade model was built for the specific problem of left ventricular contour segmentation, and the spatial and temporal features of four-chamber echocardiographic heart images were extracted. Based on the linear layer attention mechanism, the redundant information about fusion features has been reduced, and the automatic diagnosis of ventricular septal jitter abnormalities has been realized. Compared with traditional clinical diagnosis methods, the results of this study have definite advantages in diagnostic speed and accuracy, and are suitable for rapid multi-modal and multi-organ precision segmentation in clinical practice.
In the aspect of mental illness diagnosis and treatment, two works from NEU have been accepted by MICCAI2023, namely, Modeling Alzheimers Disease Progression from Multi-task and Self-supervised Learning Perspective with Brain Networks and Brain USL: Unsupervised Graph Structure Learning for Functional Brain Network Analysis. Doctoral students Liang Wei and Wen Guangqi and postgraduate Zhang Pengshuai are the first authors respectively, and Associate Professor Cao Peng and Professor Yang Jinzhu are the corresponding authors. Targeting the problems of incomplete clinical data and insufficient supervised information at the stage of diagnosis and treatment of mental diseases, this study proposed a disease diagnosis method based on unsupervised brain network structure learning and a disease progression model based on self-supervised multi-task learning. The study has been fully verified on various auxiliary diagnostic tasks for mental disorders. It has provided new insights on disease association and disease explainability, and the research results will open up a new way for clinical deployment to assist in the precision diagnosis and treatment of mental patients.
In the aspect of basic theory of lesion detection and segmentation, two works from NEU have been accepted by MICCAI2023 and IJCAI2023 respectively, namely, Lesion-aware Contrastive Learning for Diabetic Retinopathy Diagnosis and Co-training with High Confidence Pseudo Labels for Semi-supervised Medical Image Segmentation. Doctoral students Hou Qingshan and Shen Zhiqiang and postgraduate Cheng Shuai are the first authors respectively, and Associate Professor Cao Peng and Professor Yang Jinzhu are the corresponding authors. This study mainly targeted high-precision detection and segmentation tasks of low-resource lesions. A comparative learning self-supervision framework based on knowledge distillation and a collaborative mean teacher model guided by uncertainty were designed in the study. The performance of the models proposed in this study has surpassed the existing methods, and these models can be extendedly applied to multiple modal scenarios and multiple disease tasks. This study has provided valuable reference for other similar problems in the field of medical image analysis.
Additionally, NEU’s Key Laboratory of Medical Image Intelligent Computing of MOE recently published a number of academic papers in IEEE TMI, Neural Network, JBHI, AAAI and many other top international journals and conferences. This demonstrates the outstanding achievements NEU has made in innovative basic research in the field of medical image computational analysis.