脉得智能
Recently, Maide Intelligent, in collaboration with Zhongshan Hospital, Fudan University; Shanghai Tenth People's Hospital; Jinshan Hospital, Fudan University; and the Technical University of Munich, Germany, among other institutions, achieved a major breakthrough in early breast cancer diagnosis technology. Its independently developed dual-core-driven ultrasound radiomics (Dc-HR) technology, which integrates machine learning and deep learning algorithms, enables precise diagnosis of and biopsy guidance for BI-RADS category 4 breast lesions based on dual-view grayscale ultrasound images. The related research findings have been published in the internationally authoritative journal Cancer Imaging (JCR Q1). This technology is expected to substantially reduce unnecessary biopsy rates and provide a new approach to precise early diagnosis of breast diseases, particularly for primary-level medical institutions lacking high-end ultrasound equipment. Professor Xu Huixiong and Dr. Zhao Chongke of Zhongshan Hospital, Fudan University, are the corresponding authors, while Dr. Zhou Boyang of Zhongshan Hospital, Fudan University, and Tan Bin, a CV algorithm engineer at the Maide Intelligent Research Institute, are the first authors.
Breast cancer is the most prevalent malignancy among women worldwide, and grayscale ultrasound is a core modality for breast lesion screening. In clinical practice, the malignancy rate of BI-RADS category 4 breast lesions varies widely (5%-95%). Confirmation typically requires core-needle biopsy, causing many patients with benign lesions to bear unnecessary trauma, anxiety, and medical costs. Although emerging technologies such as shear wave elastography (SWE) and contrast-enhanced ultrasound (CEUS) can improve diagnostic accuracy, they depend on high-end equipment, contrast agents, and experienced physician operation, making them difficult to implement broadly at primary-level facilities. To address this clinical pain point, the research team established a dataset covering 1,054 patients and 1,062 BI-RADS category 4 lesions, completing technology development and validation over three years. The team also designed three radiomics approaches, including machine learning (ML) and deep learning (DL) single-core models, as well as the Dc-HR dual-core model integrating the strengths of both, and systematically compared their diagnostic performance and value in optimizing core-needle biopsy. The findings showed that Dc-HR technology delivered outstanding performance, with an area under the curve (AUC) of 0.944-0.980, significantly outperforming the ML model (0.544-0.882) and DL model (0.839-0.964), while also exceeding conventional enhancement technologies such as SWE (0.584-0.725) and CEUS (0.813-0.840). Supported by dual-view ultrasound images, the technology reduced the unnecessary core-needle biopsy rate to 0% in the internal validation cohort and 7.22% in the external validation cohort, while missing only 1-2 malignant lesions and maintaining an exceptionally high level of diagnostic safety. For BI-RADS category 4 lesions of different subcategories, Dc-HR technology demonstrated precise applicability: for category 4a subtypes, which have a high proportion of benign lesions, it reduced unnecessary biopsies by 90.70%-100%; for category 4b and 4c subtypes with higher malignancy risk, it maintained a positive predictive value of 96.43%-100%, enabling “precise triage.” Decision curve analysis confirmed that the technology delivered a significantly higher clinical net benefit than existing diagnostic and treatment approaches, optimizing core-needle biopsy decision-making without requiring additional equipment. “The core advantage of Dc-HR technology lies in its integration of the precise feature extraction of machine learning with the global information-capturing capabilities of deep learning. It relies solely on conventional grayscale ultrasound images, making it easy to promote at the primary level,” said the project leader. The technology can not only reduce trauma to patients with benign lesions and the waste of medical resources, but also provide reliable diagnostic and treatment tools for regions lacking high-end equipment, supporting the standardized development of early breast cancer screening and diagnosis. At present, the study has undergone multicenter validation, and the diversity of the dataset ensures the generalizability of the technology. The team will subsequently conduct larger-scale clinical trials to further optimize the model and extend it to the diagnosis and treatment of lesions across all BI-RADS categories. It will also explore multidimensional diagnostic and treatment approaches incorporating clinical information and genetic data, providing more comprehensive technical support for precision medicine in breast diseases.
Maide Intelligent Technology (Wuxi) Co., Ltd. is a high-tech enterprise deeply engaged in the field of AI healthcare, jointly established by an academician-led team and an experienced operations team. The company focuses on research into and the translational application of AI medical big data algorithms. As a Jiangsu Province Potential Unicorn Enterprise, a Specialized, Sophisticated, Distinctive and Innovative Enterprise, and an executive council member of the National Precision Medicine Industry Innovation Center, the company is committed to “empowering precision medicine with AI and building a standardized diagnostic and treatment ecosystem.” It continues to make breakthroughs in intelligent diagnosis of ultrasound imaging, developing core competitiveness particularly in the diagnosis and treatment of thyroid, breast, and other diseases, with its business covering hundreds of medical institutions at all levels. Leveraging strong proprietary AI R&D capabilities and authoritative medical resources, Maide Intelligent has established a complete closed loop of “algorithm R&D-product translation-clinical implementation.” The company has established its headquarters, regional headquarters, data centers, joint laboratories, and R&D centers in Wuxi, Shanghai, Hangzhou, Chengdu, and Hong Kong. Together with expert teams from leading Class III Grade A hospitals in China, it has developed a multi-disease AI real-time diagnosis and treatment system based on ultrasound imaging, as well as an AI bioinformatics data analysis platform for genomics. With its robust technological capabilities, the company was recognized as a winning organization in the Ministry of Industry and Information Technology's AI medical device innovation challenge, and has received multiple honors, including the Jiangsu Provincial Artificial Intelligence Science and Technology Award. In the field of breast disease diagnosis and treatment, Maide Intelligent has specifically addressed clinical pain points by establishing a full-process AI-assisted solution. Breast ultrasound examinations depend on physician experience, and issues such as insufficient consistency in BI-RADS classification interpretation, missed diagnoses of small lesions, and excessive biopsies of benign lesions are prominent, particularly at primary-level medical institutions. To this end, the company developed a dedicated AI detection model for breast nodules, enabling precise analysis and quantitative assessment of ultrasound images through deep learning algorithms. This core technology has been granted an invention patent. It can perform frame-by-frame analysis of breast ultrasound videos and effectively reduce false-positive rates through a detection-box queue screening algorithm, precisely identifying targets prone to missed diagnosis, such as microcalcification clusters and lesions with indistinct margins. As a benchmark enterprise in its market segment, Maide Intelligent has entered into in-depth strategic collaborations with leading global partners, with its products deployed in hundreds of leading and primary-level medical institutions nationwide. The company continues to deepen technological iteration in the breast field and will further optimize model generalizability, expand into the diagnosis and treatment of lesions across all BI-RADS categories, and explore multidimensional assessment approaches combining clinical information and genetic data. At the same time, relying on multicenter clinical collaboration, it will continuously accumulate real-world data, promote the deep integration of AI technology throughout the breast diagnosis and treatment process, and provide more efficient and precise intelligent healthcare support for breast health among women worldwide. Written and formatted by丨Luo Chaoran, Huang Qian Design丨Huang Ruoxi Reviewed by丨Chen Yonghong