Bridging the Digital-Physical Divide: Human-AI Collaboration as the Catalyst for Real-World AI Medical Imaging Adoption

“Humans possess unique advantages that are difficult for machines and algorithms to replicate. We have advanced motor skills that far exceed the capabilities of robots, enabling us to perform delicate manipulations. Although advances in AI and robotics are impressive, we believe that combining machine capabilities with human unique strengths will lead to higher productivity and create more value. Seizing this opportunity requires powerful human-machine interfaces to bridge the gap between the digital and physical worlds.”

In the November–December 2017 issue of Harvard Business Review, a lengthy article titled “A Manager’s Guide to Augmented Reality” was published. In this article, Michael Porter, known as the father of competitive strategy, outlined directions for integrating new technologies into various aspects of production and entrepreneurship.

The essence of medicine is evidence-based practice; therefore, AI’s application in the healthcare sector cannot achieve the same dominance as AlphaGo did on the Go board. While Go operates within a framework of rules and logic where every move has an optimal solution, medicine and disease involve vast unknowns. Consequently, AI trained on limited datasets struggles to navigate these uncharted territories of medical science.

As some experts have noted, the areas where AI is currently most effectively applied are in medical technology departments, such as pathology and radiology. While it does surpass human capabilities in certain disciplines, its overall scope within medicine remains quite limited. Even for a single disease, different subtypes pose significant challenges for AI.

Undoubtedly, the application of AI will enhance the efficiency of physicians’ work, thereby driving an overall increase in productivity. As institutions and organizations across China continue to explore new possibilities, both senior and primary-care physicians are redefining their roles and equipping themselves with emerging technologies to meet the demands of the next era.

Currently, lung cancer has the highest mortality rate among all cancers in humans.

In February 2018, the National Cancer Center released the latest national cancer statistics. The data showed that lung cancer ranked first in incidence nationwide, with approximately 781,000 new cases annually.

Early screening is a critical strategy for reducing lung cancer mortality. However, since patients with early-stage lung cancer typically lack overt clinical symptoms and specific biomarkers, the primary current screening approach relies on radiological imaging to detect suspicious pulmonary lesions.

Currently, 80%–90% of data in the healthcare industry originates from medical imaging, which serves as a critical basis for physicians’ disease diagnosis. With the exploration and application of AI in healthcare, early and precise screening for pulmonary nodules has become feasible.

On March 31, an “alternative” human-machine image interpretation experience event was held in Chengdu.

This event was hosted by the Thoracic Surgeons Branch of the Chinese Medical Doctor Association, co-organized by West China Hospital of Sichuan University, and supported with AI technology by LinkDoc Technology. It is described as “unconventional” primarily because it does not feature a direct confrontation between doctors and AI, as often depicted in previous media coverage on medical AI. Instead, the event focuses on physicians’ hands-on experience with AI-assisted image interpretation. During the session, each participating physician is required to interpret 16 cases of small pulmonary nodules within 13 minutes—eight cases with AI assistance and eight without—followed by a comparison of diagnostic outcomes between unassisted physicians and those aided by AI.

Second, this activity is not a comparative study to assess the concordance between AI and physician diagnoses (as such comparisons are susceptible to physicians’ subjective biases); rather, it uses pathological diagnosis—the clinical “gold standard”—as the benchmark for evaluation. Furthermore, based on clearly defined nodule localization, it conducts further assessment of lesion benignity or malignancy.

Third, while similar past initiatives typically relied on public datasets for model training, this event utilized raw datasets jointly provided by three hospitals. LinkDoc Technology was responsible for the collection, de-identification, conversion, and standardization of imaging data, with data annotation performed by leading domestic imaging experts. This approach ensured that participating physicians worked with novel, high-quality, and precise data, which not only accurately reflected real-time image interpretation outcomes by physicians and AI but also offers valuable insights for future clinical diagnosis and treatment practices.

Ultimately, the results of this human-AI collaborative image interpretation experience were not surprising; the diagnostic outcomes from doctors + AI were significantly superior to those from doctors alone. (For detailed results, please refer to:128 Imaging Cases, 32 Medical Experts Explore “Human-Machine Collaboration”—AI Image Interpretation Takes Center Stage at the 2018 Annual Meeting of Thoracic Surgeons）。

This seems to send a signal: at this stage, “human-machine collaboration” may be the key to whether doctors and AI can successfully enjoy their “honeymoon.”

On-site expert commentary further corroborates this view: although thoracic surgeons already achieve high diagnostic accuracy, AI can still further enhance both the accuracy and speed of diagnosing small pulmonary nodules. Moreover, the application of AI-assisted diagnosis for small pulmonary nodules holds particular significance for primary care hospitals, where rates of misdiagnosis and missed diagnosis are relatively high.

Of course, as the closest counterpart to radiology, pathology is likewise an ideal scenario for AI implementation.

Take pathologists as an example; this is a group that is severely lacking among doctors in China. The National Health Commission's Statistical Yearbook shows that the total shortage of pathologists in China can reach 100,000. Even for well-trained pathologists, there are differences in their diagnoses for the same patient, and these differences are an important cause of misdiagnosis. For instance, the consistency rate among doctors in diagnosing certain forms of breast cancer and prostate cancer is as low as 48%.

To make an accurate diagnosis, physicians must evaluate a vast amount of diagnostic information. Typically, pathologists are responsible for reviewing whole-slide images comprising over 1,000 megapixels, with accountability for every pixel, which entails processing substantial volumes of data. Consequently, doctors face significant time constraints. To achieve high diagnostic accuracy within limited timeframes, the integration of AI into digital pathology research has emerged as a viable solution.

AI can not only shorten the time required for pathological diagnosis and improve diagnostic efficiency, but also provide more accurate diagnostic results. The effective use of AI can truly help pathologists enhance their interpretive skills, starting with precise diagnosis to genuinely realize precision medicine.

Certainly, beyond imaging and pathology, AI companies are also making their presence felt in areas such as medical record/literature analysis, virtual assistants, and new drug development.

In addition to clear application scenarios, policy support is equally essential for the integrated development of AI and physicians.

On December 14, 2017, the Ministry of Industry and Information Technology issued the “Three-Year Action Plan for Promoting the Development of New-Generation Artificial Intelligence Industry (2018–2020),” which pointed out that a new round of scientific and technological revolution and industrial transformation is currently emerging. The formation of big data, innovations in theoretical algorithms, enhancements in computing power, and advancements in network infrastructure are driving artificial intelligence into a new stage of development, with intelligence becoming a key direction for technological and industrial progress.

The notice sets specific targets for the currently popular AI-assisted medical imaging diagnostic systems: to promote the standardization and normalization of medical image data acquisition; to support the research and development of AI-assisted diagnostic technologies for medical imaging in typical disease areas, including the brain, lungs, eyes, bones, cardiovascular and cerebrovascular systems, and breast; and to accelerate the productization and clinical auxiliary application of AI-assisted medical imaging diagnostic systems. By 2020, domestically advanced multimodal AI-assisted medical imaging diagnostic systems are to achieve a detection rate of over 95%, a false-negative rate of less than 1%, and a false-positive rate of less than 5% for the aforementioned typical diseases.

From the perspective of calculation formulas, false negatives, true negatives, true positives, and false positives—these four metrics can be used to derive the sensitivity, specificity, negative predictive value, and positive predictive value of a diagnostic method. Sensitivity, also known as the true positive rate, reflects the missed diagnosis rate of a method in identifying a specific lesion. Specificity, also known as the true negative rate, reflects the misdiagnosis rate of a method in identifying a specific lesion.

Currently, the system sensitivity in the medical AI industry generally exceeds 90%. This is because if a system has poor sensitivity and fails to detect suspected nodules, physicians may bear corresponding liability. However, it is challenging to balance both specificity and sensitivity; therefore, many companies prioritize sensitivity during system development, aiming to improve specificity while ensuring high sensitivity.

In addition to policy support for promoting the development of the artificial intelligence industry at the national level, legal and regulatory issues involved in the application process of AI also require early planning and oversight. Particularly in the heavily regulated healthcare sector, there are still many issues regarding the commercial application of artificial intelligence that need to be standardized through policy.

First, standards for the application of artificial intelligence. Medical issues involve human health and life, constituting a complex and sensitive domain where every matter is closely tied to patient safety. Therefore, the industry urgently needs clear regulatory measures at the national level, using legislation to govern factors such as the scope of AI applications in healthcare, the extent of regulatory oversight, and the determination of liability for risks.

Second, the reasonable and lawful application of data. Because artificial intelligence needs to learn and iterate from past data through technologies such as natural language recognition, it can possess intelligence and improve. Therefore, a large amount of high-quality medical data based on the real world will become the foundation and guarantee of the accuracy of artificial intelligence.

Although national policies and plans continuously encourage the application of artificial intelligence in the healthcare industry, physicians’ resistance to AI remains a significant challenge at present. Doctors are concerned that AI may become a tool to replace them. In light of this situation, human-AI collaboration has emerged as the optimal solution.

The significance of human-AI collaboration lies in ensuring that artificial intelligence does not become the “culprit” replacing humans, but rather allows physicians to perceive it as a tool akin to a computer.

For physicians at hospitals of different tiers, the application value varies significantly. For primary care physicians, human-AI collaboration can substantially reduce misdiagnosis rates and improve diagnostic and treatment standards; for specialists, it enhances clinical efficiency, truly returning time to physicians.

Policy is the foundation. So, where exactly are the application scenarios for AI?

Currently, to alleviate the strain on medical resources and address the uneven distribution of high-quality healthcare services, China is vigorously promoting tiered diagnosis and treatment systems as well as medical consortia. The aim is, under the core guidelines of tiered diagnosis and treatment, to channel high-quality medical resources from top-tier hospitals down to the grassroots level, thereby achieving a 90% consultation rate within county-level jurisdictions—commonly referred to as “keeping major diseases within the county.”

However, practical implementation poses significant challenges, as physicians at large hospitals have inherently limited availability. The concept of addressing regional disparities through telemedicine is similarly constrained by physicians’ time limitations. Therefore, the scarcity of high-quality physician resources remains the core issue in primary healthcare.

Although medical consortiums can facilitate coordination between tertiary and primary hospitals to some extent, it is difficult to significantly improve the clinical competence of grassroots physicians in a short period through measures such as specialty support or remote consultations. The cultivation of healthcare professionals remains a long-term endeavor.

This situation has enabled AI to generate application value at the grassroots level.

“Human-AI Collaboration” between clinicians and AI can match or even surpass the performance of industry experts in terms of time efficiency and accuracy, leveraging clinical data from large hospitals. For primary care physicians, using AI is akin to having a top-tier specialist provide personalized, on-the-job training.

Particularly in high-incidence severe diseases such as pulmonary nodules, breast cancer, and cervical cancer, AI assists primary care physicians in patient screening with high efficiency and accuracy. With only minimal training, it can empower primary care physicians to deliver diagnostic and therapeutic care at the level of specialists.

Moreover, there is no conflict between primary care physicians leveraging AI to enhance their diagnostic and treatment capabilities and the practice of telemedicine.

Due to the differing roles and functions of hospitals at various levels, primary care physicians mainly treat common and chronic diseases, while specialists at higher-level hospitals focus on treating specialized conditions. Therefore, the trigger scenario for telemedicine should be the demand for specialized care arising from the geographical separation between patients and hospitals. Leveraging its professional capabilities, AI can assist primary care physicians in screening patients based on disease categories, thereby facilitating triage and referral.

The essence of technology ultimately lies in serving humanity. Whether it is artificial intelligence, big data, or informatization, these technologies aim to address tasks that are difficult for physicians to accomplish due to constraints on time and energy. Taking big data as an example, what physicians need is not merely the data itself, but the insights derived from it. Therefore, the true value of the current big data industry resides in data quality, data processing and analysis, and the extraction of data-driven insights.

This is precisely the significance of AI for physicians. It liberates them from routine, labor-intensive tasks, allowing them to devote more energy to medical research and humanistic care. Their research findings are then learned by AI, which in turn feeds back into their practice, further enhancing diagnostic and therapeutic efficiency.