Zhejiang University Ruiyi AI Medical Research Center Files IPO Prospectus: 3 MICCAI 2019 Papers and Clinical Deployment in 14 Hospitals

Zhejiang University Ruiyi Artificial Intelligence Research Center (a university-level research center of Zhejiang University) has continuously participated in MICCAI submissions since its establishment. As of the conclusion of the current submission cycle, six manuscripts from the Zhejiang University Ruiyi Artificial Intelligence Research Center have been accepted by MICCAI.

In 2019, the Zhejiang University Ruiai Artificial Intelligence Research Center had a total of three papers accepted by MICCAI. These studies covered research directions such as object detection, semi-supervised object segmentation, and multimodal recognition, achieving innovative breakthroughs in the application of artificial intelligence in healthcare. Meanwhile, the outcomes of cervical lesion identification based on multimodal fusion have been implemented in 14 hospitals, including the Women’s Hospital School of Medicine Zhejiang University, Wenzhou People’s Hospital, the First Affiliated Hospital School of Medicine Zhejiang University, the Second Affiliated Hospital School of Medicine Zhejiang University, the Affiliated Hospital of Qingdao University, Shandong Provincial Third Hospital, and Wuhan Tongji Hospital, demonstrating favorable results in clinical applications.

Below is an overview of the research achievements of the three selected papers:

1.Jintai Chen,  Yanjie Wang, Ruoqian Guo, Bohan Yu, Tingting Chen, Wenzhe Wang, Ruiwei Feng, Danny Z. Chen, Jian Wu. LSRC: A Long-Short Range Context-Fusing Framework for Automatic 3D Vertebra Localization. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2019.

CT-based spinal bone localization plays a significant role in clinical diagnosis and treatment; however, its performance is often suboptimal due to complex pathological conditions. In this paper, we propose a network framework termed LSRC, which fully integrates local and long-range contextual information. Specifically, our method leverages both 3D and 2D CT data to fuse local and long-range features, and employs a novel attention-based optimization module to refine the localization results. Our approach outperforms state-of-the-art methods on public spinal CT datasets encompassing various pathological conditions.

(Network architecture diagram. The 3D and 2D networks are used to extract local and long-range information, respectively, and the fused features are refined by the Global Refinement Module (GRM).)

2. Tingting Chen, Xinjun Ma, Xuechen Liu, Wenzhe Wang, RuiweiFeng, Jintai Chen, Chunnv Yuan, Weiguo Lu, Danny Z. Chen, Jian Wu. Multi-View Learning with Feature Level Fusion for Cervical Dysplasia Diagnosis. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2019.

The primary contribution of this paper is the proposal of an innovative multimodal fusion model that addresses the limitation of existing fusion methods in fully exploring latent correlations between different modalities, achieving state-of-the-art performance in cervical lesion recognition on a colposcopy dataset. We propose a Feature-Level Fusion architecture that integrates features during the extraction phase from acetic acid images and Lugol’s iodine images, thereby better leveraging their respective features and learning the inter-modal relationships. Furthermore, we investigate two fusion strategies—unidirectional and bidirectional fusion—based on the direction of information flow. Experimental results demonstrate that bidirectional fusion enables mutual enhancement between the two modalities during feature learning, yielding superior performance.

(Network architecture diagram: unidirectional fusion (left) and bidirectional fusion (right); the model backbone is ResNet50, and the Assistant Modules represent the specific fusion structures)

3. Han Zheng, Lanfen Lin, Hongjie Hu, Qiaowei Zhang, Qingqing Chen, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, Ruofeng Tong, and Jian Wu. Semi-supervised Segmentation of Liver Using Adversarial Learning with Deep Atlas Prior. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2019.

Compared to natural images, medical images present greater research challenges due to their limited data volume and the difficulty of annotation. However, medical images possess anatomical priors absent in natural images, such as the shape and location of organs. Incorporating anatomical prior knowledge into the training of deep learning models is an effective approach to improving medical image segmentation performance. Based on the probabilistic atlas, an anatomical prior for medical images, this paper proposes the DAP (Deep Atlas Prior) loss. This loss function assists neural network training by directing the model’s attention to challenging regions, particularly organ boundaries. Furthermore, traditional losses such as cross-entropy loss and focal loss are treated as likelihood losses; combined with the proposed DAP loss, a Bayesian loss is further formulated within the framework of Bayesian models. The model employs a Generative Adversarial Network (GAN) for semi-supervised learning, incorporating unlabeled data into the training phase to alleviate the burden of data annotation. The proposed method achieves superior segmentation performance on liver and spleen datasets, thereby enhancing the model’s analytical and processing capabilities for medical images.

(Overall Network Architecture Diagram: Semi-Supervised Adversarial Network with Bayesian Loss)

The primary objective of the Ruiyi AI Research Center at Zhejiang University is to establish an open, collaborative innovation platform integrating industry, academia, and research. By leveraging the technological strengths in medical artificial intelligence (AI) from Zhejiang University’s College of Computer Science and Technology, College of Information Science and Electronic Engineering, School of Medicine, College of Pharmaceutical Sciences, and affiliated hospitals, the Center aims to create China’s own open medical AI platform. This initiative will support data mining, application, and security assurance in the healthcare sector. Through medical big data computing and AI, the Center will enable functions such as assisted diagnosis and treatment and precise doctor-patient matching, thereby improving physicians’ work efficiency and providing patients with personalized, precision healthcare experiences, ultimately driving the development of the medical AI industry.