Designing the World's First AI-Powered Clinical Trial in Thoracic Surgery: An Interview with Dr. Chen Xiuyuan of Peking University People's Hospital

If we consider surgeryAnalogyCheng drives, soA Map of the Human Body Is Indispensable。

On the chest X-rayTimes，ThisMap is aRoughText description, informWeLeft LungHereThere is aThings，ShouldSurgical AccessExploration。

On CTThe Era, maphas become much more precise,It has become a paper map, showing you what Beijing looks like, and thenPeople's Hospital"There's a tumor here; you should get it checked out."

During the AI 3D reconstruction phase, itEvolved Again,Transform into an electronic map。We can zoom in and out freely.、Rotate it to align with the direction of the path currently in my line of sight.

With a goodMapOnly then can subsequent interventions beFurtherStrengthen. 3D Reconstruction of CTFor surgery, it is moving towards automationofA significant step, akin to the transition from paper maps to electronic maps.

Dr. Chen Xiuyuan said so.

Prior to the widespread adoption of three-dimensional lung reconstruction,Dr. Chen XiuyuanThoracic surgeons, represented by [specific individuals/groups], often face numerous challenges.

The inherent complexity of pulmonary anatomy and the significant mobility of organs during surgery create substantial challenges in image interpretation. Taking the dorsal segmental artery as an example, it is difficult to determine solely through preoperative imaging whether it is a single vessel or bifurcated, how it behaves under surgical traction, and whether this impacts the procedure. Such determinations often rely on real-time intraoperative judgment.

This complexity restricts young surgeons to learning within limited operative time and with a small number of cases, significantly compressing opportunities for experience accumulation and prolonging the learning curve. Consequently, variations in experience lead to cognitive biases, resulting in substantial inter-surgeon variability in the interpretation of CT images.

Dr. Chen Xiuyuan remarked with a smile that many understandings of pulmonary anatomy, accumulated since his student days, were later proven incorrect through AI-assisted intelligent 3D reconstruction.

In 2020, with the gradual refinement of 3D lung reconstruction algorithms within the industry, the possibility of clinical application was no longer a fantasy. Dr. Chen Xiuyuan began to consider how to design clinical trials.

Randomized controlled trials (RCTs) are a major type of clinical trial, but they may not be suitable for AI imaging technologies, particularly in the future envisioned by Dr. Chen Xiuyuan.

If an RCT study design follows the framework of a Phase III clinical trial, treating AI-based 3D reconstruction as an intervention, the selected endpoints must be measures of perioperative efficacy and safety. However, significant challenges exist in terms of data reliability and trial duration, whether the endpoints are perioperative bleeding or the rate of postoperative reoperation, or longer-term oncological survival outcomes. In particular, achieving a sufficient number of positive events may require a sample size of tens of thousands of patients.

“A sample size in the tens of thousands is enough to make any surgical clinical researcher furrow their brow.”

The sheer scale of such clinical trials presents formidable challenges in patient enrollment and exclusion, timeline management, research costs, and physician acceptance of AI.

Dr. Chen Xiuyuan needs to find clues from “AI” and “imaging.”

Multi-Reader Multi-Case (MRMC) design is aFor Assessing the Diagnostic Accuracy of Medical Imagesmedical research methodology. This design evaluates the performance of multiple physicians (readers) in completing a specific medical task across multiple cases. Through this approach, we can assess inter-reader agreement and compare the efficacy of different testing methods or imaging techniques.

In MRMC study designs, each physician reviews a series of distinct cases and provides diagnoses based on their professional judgment. Researchers then analyze these diagnostic outcomes to evaluate physician performance across varying scenarios. The advantage of this design lies in its simultaneous consideration of both reader and case variability, thereby better reflecting real-world clinical practice.

Dr. Chen Xiuyuan stated that MRMC can be colloquially understood as a study design situated between diagnostic trials and interventional trials. On one hand, the output generated by physicians after reviewing imaging cases represents a form of surgical cognition; however, this cognition is not a binary variable in the traditional diagnostic sense, but rather a more complex and comprehensive assessment. On the other hand, while this cognition influences surgical decision-making to some extent, the trial is conducted under idealized conditions and does not directly impact the enrolled patients. MRMC studies evaluate the triangular relationship among the new method, physicians, and patients. This additional layer involving physicians is particularly applicable to AI-assisted medical imaging technologies.

After nearly two years of preliminary preparation and repeated refinement, in 2022Dr. Chen XiuyuanThe thoracic surgery team at Peking University People’s Hospital collaborated with Infervision to conduct a multicenter clinical trial.

The trial enrolled 400 patients at Peking University People's Hospital, Shanghai Pulmonary Hospital, and the Second Xiangya Hospital. Subsequently, data meeting the required sample size were randomly selected from these cohorts to constitute the final dataset for analysis.

From a Surgical Perspective, the dataset included patients who underwent lobectomy and segmentectomy, with lesions distributed across all five lung lobes.From the Perspective of Preoperative CT Image Parameters, the trial covered four CT equipment manufacturers with high market penetration (GE, Philips, Siemens, and United Imaging), further ensuring the representativeness of the study data.

Dr. Chen Xiuyuan has multiple considerations regarding the design of this trial.

First, patients were recruited from representative centers with high outpatient volumes and broad geographic distribution, ensuring comprehensive coverage. Furthermore, this clinical trial innovatively selected accuracy in anatomical structure identification and time efficiency as its clinical endpoints.Most importantly, the rigor of the study is enhanced through the MRMC fully crossed design.

The so-called fully crossed design refers to a setup in which all readers provide interpretations for the imaging data of all cases under each reading modality. To account for potential bias arising from the order of AI-assisted reading, the participating radiologists were randomly assigned to two groups, and the trial was conducted in two phases. In the first phase, one group of radiologists performed readings with AI assistance, while the other group performed readings without AI assistance. After a 28-day washout period, the two groups switched to the alternative reading method in the second phase.

By adopting this most rigorous MRMC design, Dr. Chen Xiuyuan and Infervision were able to minimize the impact of inter-reader variability through each radiologist’s self-controlled before-and-after comparison, thereby accurately reflecting the quantifiable assistance that AI-based 3D reconstruction technology provides to physicians across different levels of seniority.

Leveraging this opportunityThe World's First AI Clinical Trial in Thoracic Surgery, TuiXiang Medical’s thoracic surgical planning product has successfully obtained Class III certification, making it the first and only lung surgery planning product approved with a Class III certificate by the NMPA. It can serve as a reference for thoracic surgeons in formulating lung surgery plans. The trial met its primary endpoint, demonstrating that the AI system can effectively enhance physicians’Anatomical Structure Recognition Accuracy，Accuracy of Surgical Procedure SelectionandSurgical Resection MarginAccuracy of Determination。

The surgical planning product has received Class III medical device certification, delivering dual value to both patients and physicians.From the patient’s perspective, 3D reconstructed images can enhance their understanding of the surgical procedure, alleviate fear, and boost confidence. Meanwhile, as AI-based 3D reconstruction reduces reliance on contrast agents, it can effectively mitigate contrast-related safety risks and lessen the financial burden for patients who do not require contrast-enhanced chest CT scans. From the physician’s perspective, AI-based 3D reconstruction holds significant importance in improving surgical certainty, reducing procedural difficulty, increasing success rates, and ensuring compliance with hospital admission regulations.

ForAIClinical Applications,Dr. Chen XiuyuanThere are three layers of perception: near, middle, and far.

At the end closest to oneself in application, it isThe Impact of AI-Assisted 3D Reconstruction on Medical Education and Doctor-Patient Communication.As previously mentioned, 3D reconstruction has significantly enhanced the understanding of pulmonary anatomy among physicians, including Dr. Chen Xiuyuan, and has substantially shortened the learning curve for subsequent trainees. In terms of doctor-patient communication, Dr. Chen expressed relief that he no longer needs to use his five fingers as an analogy for lung locations; instead, he can now leverage 3D visualization to precisely convey the location and size of the resection area to patients.

In the foreseeable future, Dr. Chen XiuyuanWe hope to see AI applications transition from diagnosis to intervention.From relatively soft interventions (such as imaging-based risk alerts and assistive planning of surgical approaches) to more robust interventions (such as direct operative assistance and robotic correction of erroneous movements), with humans ultimately calibrating and correcting the machine’s actions.

And with the advent of AI, regarding the long-term changes in the healthcare sector itself,Dr. Chen XiuyuanOffers a unique interpretation.

Dr. Chen Xiuyuan believes that with the advent of medical AI, the problems that AI-driven products and technologies aim to address and are capable of addressing have become increasingly diversified. Are traditional endpoint indicators still absolutely applicable? Do metrics such as physician and patient confidence and satisfaction truly hold value? More factors require comprehensive consideration; as the dimensions of endpoint indicators expand, how should researchers strike a balance from a statistical perspective?

These insights from clinicians on AI have undoubtedly provided fresh perspectives for both the research and industrial sectors. It is precisely through these insights that forward-looking trial designs, better aligned with the demands of new technologies and products, have been made possible.