At the 2024 Global Digital Economy Conference, visitors visited AI medical equipment used to assist doctors in imaging diagnosis. Chen Xiaogen
A single CT scan can help doctors identify a variety of cancers, and the online platform can complete the connection of personalized medical resources in seconds... In recent years, artificial intelligence (AI) technology is comprehensively revolutionizing all aspects of tumor diagnosis and treatment.
"AI can run through the entire process of tumor diagnosis and treatment." Li Zhicheng, executive director of the Medical Artificial Intelligence Research Center of the Institute of Biomedicine and Health Engineering, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, told a reporter from Science and Technology Daily, "From initial imaging diagnosis, lesion identification, patient admission, to Pathological diagnosis, visualization of surgical plans, and even discharge recovery tracking, the intervention of AI is visible and tangible to doctors and patients.”
Help early screening of tumors
Xu Zhonghuang, director of Beijing US-China Airui Cancer Hospital, said that many cancer patients are already in the mid-to-late stage when diagnosed and miss the best opportunity for treatment. Early screening can help doctors detect conditions at the asymptomatic or precancerous stage, and effectively reduce morbidity and mortality through early intervention. AI has great potential in the field of early tumor screening.
Early tumor screening usually relies on a series of non-invasive or minimally invasive examination methods, including imaging examinations, blood marker detection and molecular diagnosis. In this regard, AI intervention has made breakthrough progress. Li Zhicheng believes that with the support of image-based deep learning technology, AI's performance in certain tumor screenings can even surpass that of human experts.
In the past two years, international journals such as Nature have published multiple studies on AI-assisted tumor screening. The CHIEF model developed by the Harvard Medical School team can not only diagnose 19 types of cancer, but also locate the tumor microenvironment, guide treatment strategies and predict survival rates. The pancreatic cancer early detection model PANDA developed by Alibaba Damo Academy has an accuracy of 92.9% in determining the presence of lesions. These results show that AI can not only assist diagnosis, but also play a key role in precise treatment.
Related practices have shown the role of AI in tumor screening. In February this year, Alibaba's "Medical AI Multi-Cancer Early Screening Charity Project" was deployed in Lishui Central Hospital and other institutions in Zhejiang, applying the innovative medical AI technology of DAMO Academy to the health field. "The project screened more than 50,000 people within 4 months. The screened diseases included pancreatic cancer, esophageal cancer, gastric cancer, and colorectal cancer. 145 cancer lesions found among them have been clinically confirmed." Damo Academy Medical AI Team The person in charge, Lu Le, explained that by combining a large amount of historical data and complex algorithms, AI can extract information about tiny lesions that are difficult to detect with the naked eye from images. In tedious image analysis tasks, AI can also quickly process large amounts of data, reducing the pressure on doctors.
Xu Zhonghuang said that cancer must rely on multidisciplinary collaboration to formulate optimal treatment plans, and AI can help solve problems such as shortage of professionals and high economic costs in this process.
Taking PANDA as an example, Lu Le said that the model is equivalent to bringing together the knowledge base of dozens of doctors from different professions, and achieves cross-department data fusion by integrating multi-modal data such as imaging data, genomic information, pathology data, etc. On this basis, the model can extract key lesion information and potential pathological characteristics, and then carry out comprehensive analysis across departments.
Improve cancer awareness
Promoting scientific understanding in the medical field is a higher dimension for AI to assist tumor diagnosis and treatment.
Li Zhicheng’s team has been engaged in glioma research for decades. Talking about the current status of diagnosis and treatment of glioma, Li Zhicheng said: "Our scientific understanding of this disease is still limited. Doctors have not yet fully understood the occurrence, development and recurrence mechanism of glioma, and have not yet found effective and precise treatments. way."
Xu Zhonghuang feels the same way. "The lack of knowledge about cancer limits the methods of diagnosis and treatment. In the face of difficult and complicated diseases, many times in clinical practice we can only cross the river by feeling the stones."
Existing AI diagnosis and treatment models also have limitations. Li Zhicheng said that many models are trained through large-scale annotation data sets to find correlations between image features and clinical outcomes. Although this method has achieved remarkable results in terms of accuracy, this "black box" operation lacks explanatory basis, making it difficult for doctors to fully trust the AI's diagnostic results. Therefore, it is particularly important to return to the source of medical knowledge.
In this regard, AI has a lot of room to play. "AI can integrate multi-modal data such as imaging, pathology, genes, etc., provide multi-scale comprehensive analysis, and help us build a more complete 'portrait' of tumors. Tumor is an ecosystem composed of complex cancer cells, and the more detailed its portrait is. The more accurate it is, the more it can discover tumor behaviors and potential treatment targets that have been ignored in the past, providing new ideas for front-end treatment. "Li Zhicheng said that with the continuous enrichment of molecular-level data such as genomes and proteomes, AI is expected to break through the existing cognitive bottlenecks. Help improve scientific understanding of complex cancers.
Xu Zhonghuang added: "In the face of unfamiliar tumors, if AI can advance human understanding of them, even a small step, it may fundamentally provide new methodological guidance for tumor diagnosis and treatment and truly change the way we deal with cancer."
Give full play to the role of data as "nourishing"
In order for AI to further empower the entire process of tumor diagnosis and treatment, it is crucial to obtain high-quality, comprehensive, and huge data support.
The training of AI models not only relies on doctor annotations, but also requires complete clinical cycle data. Lu Le gave an example: "During the PANDA model training process, doctors not only need to provide multi-modal data such as pathological pictures, pathology reports, and CT images, but also need to manually confirm the location of the lesion and accurately outline it on the enhanced CT. Then, the engineer passes The three-dimensional image registration technology maps the three-dimensional outline of the lesion onto the plain CT image, and ultimately allows the AI to learn to identify the appearance of early pancreatic tumors in the plain CT image.”
In this process, only doctors and the AI team work closely together to provide high-quality training data for the model. Lu Le further explained that cutting-edge medical AI algorithm teams often rely on a wide range of cooperative hospitals to provide diverse data, which is crucial to improving the generalization ability of the model. Data from different hospitals provides the AI model with rich pathological background, helping it respond to various clinical scenarios more accurately.
However, due to problems such as large amounts of data required, many departments involved, and scattered data, data acquisition has become the main bottleneck in current cancer AI research. "It is not difficult to obtain a single image or pathology data, but it is very difficult to obtain all-modality data such as imaging, pathology, and genes for the same patient at the same time." Li Zhicheng said that this not only requires close cooperation among multiple departments, but also takes a lot of time. Current cancer research is often scattered among different disciplines, with image analysis being handled by imaging and engineering technicians, while genetic data is processed by molecular pathology or bioinformatics personnel. Breaking down barriers between disciplines and integrating data remains a huge challenge.
"Data is the basic 'nutrient' for whether AI can fully play its role in medical care." In Xu Zhonghuang's view, the scalability, standardization and security of data are key considerations for hospitals when deploying medical AI. Hospitals must start from the present when planning their AI layout, ensure standardization of data entry, archiving and management, design a reasonable data management framework in advance, and reserve interfaces for future data processing. The advantage of AI is that it can continuously absorb new data and optimize itself. This requires the hospital's data storage system to be scalable to cope with the growing demand for multi-modal data.
In terms of data security, Xu Zhonghuang believes that hospitals need to establish strict data encryption and privacy protection mechanisms to ensure that technology applications can provide reliable support for clinical diagnosis and treatment under the premise of complying with laws, regulations and social ethics.