Tencent Miying Showcased at Asian Society of Cardiovascular and Thoracic Surgeons Annual Meeting: AI as a Transformative Force in Medical Imaging Management

In recent years, AI has gradually entered the medical field. Artificial intelligence continues to break through the sensitivity and specificity of machine-assisted diagnosis at the “digital” level, demonstrating its value across multiple scenarios. However, many products remain some distance away from large-scale clinical application, and gaining recognition from clinicians is no easy feat. For leading experts in the medical community, the key question is whether AI’s capabilities truly deliver clinical value.

VCBeat (WeChat ID: vcbeat) has learned that the 27th Annual Meeting of the Asian Society for Cardiovascular and Thoracic Surgery (ASCVTS) was held in Chennai, India, from February 21 to 24, 2019. Thoracic and cardiovascular surgeons from around the world attended this premier event, representing the highest level of medical expertise in the field across Asia, to engage in in-depth exchanges and discussions on the latest advancements, clinical experiences, and basic research in thoracic and cardiovascular surgery. Among the key topics discussed was the value of artificial intelligence (AI) in clinical diagnosis. A prospective small-sample study conducted by Professor Zhang Lanjun’s team from the Department of Thoracic Surgery at Sun Yat-sen University Cancer Center, in collaboration with “Tencent Miying,” drew significant attention at the conference. The team was invited to deliver a presentation, and their findings affirmed the value of AI diagnostic systems in the early detection of pulmonary nodules.

Lung cancer is the most common malignant tumor worldwide, with both its incidence and mortality rates ranking first among all malignancies, making it a recognized killer of human health. The prognosis of lung cancer is closely related to its clinical stage. Due to the late onset of symptoms and signs, most patients have already developed metastasis at their initial medical consultation, missing the optimal window for surgical intervention and resulting in a 5-year survival rate of only 16%. In contrast, the 5-year survival rate for patients with Stage I disease can exceed 70–90%. Early detection of lung cancer can significantly improve patient prognosis.

Therefore, establishing a rational and effective screening protocol to conduct simple yet efficient screening for high-risk populations is a key focus of clinical practice. Clinicians are also continuously seeking suitable methods for lung cancer screening among emerging and more sensitive imaging technologies.

In August 2002, the U.S. National Lung Screening Trial (NLST) initiated a randomized controlled clinical trial comparing low-dose spiral computed tomography (LDCT) with conventional chest radiography for lung cancer screening. This study remains the most authoritative and highest-level evidence-based lung cancer screening research conducted globally to date. The prospective study demonstrated that lung cancer screening using LDCT doubled the diagnosis rate of stage I lung cancer and reduced lung cancer-specific mortality by 20%. Consequently, based on these findings, this screening modality has been recommended by multiple international authoritative guidelines and expert consensus statements for early lung cancer screening.

However, the NLST study also found that among patients in whom pulmonary nodules were detected through clinical screening with low-dose spiral CT, only 0.6–2.7% were ultimately diagnosed with lung cancer. This indicates that, given the current feasibility of early detection of pulmonary nodules via CT screening, the primary challenge facing clinicians today is how to improve the early diagnostic rate of lung cancer among patients with pulmonary nodules.

In traditional methods for the early diagnosis of pulmonary nodules, reliance on imaging alone necessitates long-term radiological follow-up to monitor morphological changes, thereby posing risks of potential radiation-induced harm. Invasive diagnostic procedures, or even direct surgical intervention, not only cause physical and psychological distress to patients but also exacerbate the burden on China’s healthcare expenditures and lead to unnecessary waste. However, the rapid advancements in novel liquid biopsy techniques and artificial intelligence–based diagnostics have ushered in a revolutionary dawn for the early detection of pulmonary nodules.

As one of the largest oncology centers in China, integrating medical care, teaching, research, and prevention with the strongest academic strength, Sun Yat-sen University Cancer Center ranks at the forefront nationwide in terms of disciplinary status and comprehensive strength. It is also continuously exploring more rational and effective lung cancer screening protocols. As early as 2017, Sun Yat-sen University Cancer Center began collaborating with Tencent to pilot the application of its AI-based medical imaging product, “Tencent Miying,” in the lung cancer screening process.

It is reported that Tencent Miying has leveraged AI-based medical image analysis to assist physicians in screening for esophageal cancer, pulmonary nodules, diabetic retinopathy, colorectal tumors, breast cancer, and other conditions. Additionally, its AI-assisted diagnostic engine helps clinicians identify and predict risks associated with over 700 diseases. In the detection of pulmonary nodules, Tencent Miying employs computer vision and deep learning technologies to aid radiologists in image interpretation through its AI-powered medical image analysis capabilities. It can precisely localize minute pulmonary nodules larger than 3 mm, achieving a sensitivity of 85% and a specificity of up to 90% in differentiating between benign and malignant lesions.

Dr. Yu Xiangyang, a student of Professor Zhang Lanjun, shared his research on AI-assisted physician diagnosis at the conference. The study prospectively enrolled 100 patients with pulmonary nodules detected by low-dose spiral CT, whose nodules were deemed to require biopsy or surgery according to the Fleischner Society’s “Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images” (2017 edition). Subsequently, the imaging data of all these patients were independently reviewed by a chief physician specializing in thoracic radiology, while simultaneously being processed by the Tencent Miying system for intelligent diagnosis. Finally, all patients underwent surgical resection of their pulmonary nodules, yielding definitive paraffin-embedded pathological diagnoses. In the final comparative analysis, Tencent Miying achieved an independent diagnostic accuracy of 79% for pulmonary nodules, surpassing the accuracy of manual diagnosis by the thoracic radiology specialist.

Professor Zhang Lanjun provided a simple and easy-to-understand example: Relying solely on photographs, it is difficult to accurately distinguish between transgender women and cisgender women. This is because transgender women can undergo bone contouring surgery to make their frontal bones, zygomatic bones, Adam’s apples, and other features indistinguishable from those of cisgender women, thereby leaving minimal identifiable characteristics in photographs. However, sex chromosome testing can readily resolve this diagnostic challenge. By the same token, in standalone CT images, the characteristic differences between early-stage lung cancers originating from the terminal bronchioles or alveolar walls and benign pulmonary nodules are inherently limited for AI acquisition and deep learning. This limitation has led to inconsistent results in current studies using multiple AI-assisted diagnostic systems; while AI has demonstrated higher accuracy than radiologists in differentiating between benign and malignant pulmonary nodules, it exhibits instability in open and uncertain environments.

How can we further leverage artificial intelligence to enhance diagnostic performance? Current liquid biopsy technologies can detect trace amounts of biological markers released into the bloodstream by early-stage tumors, such as microRNA, circulating tumor DNA (ctDNA), and circulating tumor cells (CTCs). Professor Zhang Lanjun believes that combining liquid biopsy biomarkers with artificial intelligence will significantly improve the accuracy of early pulmonary nodule diagnosis. Indeed, in the ABC model integrating clinical features (Clinic), biomarkers (Biomarkers), and AI results (AI), the area under the curve (AUC)—a statistical measure of diagnostic performance where values closer to 1 indicate higher efficacy—reached as high as 0.955. In subsequent validation cohorts, the ABC model also demonstrated higher AUC and sensitivity compared to other models, indicating that the diagnostic model combining biomarkers and AI offers greater accuracy.

“Therefore, we believe that at a time when biomedical imaging technology is not yet mature, constructing a multimodal mathematical model integrating ‘biomarkers + AI’ represents the ideal paradigm for the current application of artificial intelligence in the clinical diagnosis of pulmonary nodules,” stated Dr. Yu Xiangyang at the Third Huaxia Medical Thoracic Surgery Forum. He noted that the translation and application of AI diagnostics are currently predominantly led by basic medical schools and engineering colleges, resulting in a disconnect from actual clinical needs and subsequent optimization recommendations. There is an urgent need for clinically driven translational research spearheaded by clinicians.

At the Annual Meeting of the Asian Society for Thoracic and Cardiovascular Surgery, Professor Zhang Lanjun shared with thoracic and cardiovascular surgeons from around the world the findings of a prospective small-sample study conducted by his expert team in collaboration with Tencent Miying. He affirmed the value of artificial intelligence diagnostic systems in the early detection of pulmonary nodules, noting that:

Artificial intelligence represents a monumental transformation in the classification and management of traditional imaging data, capable of rapidly and simultaneously processing tens of thousands of images. This will significantly reduce the physical and cognitive burden on highly skilled radiologists. Tencent Miying, built on state-of-the-art deep convolutional neural network algorithms, can be directly deployed or adapted for machine deep learning across hospitals of varying tiers, thereby reducing the sample sizes required for model adaptation at different institutions. With the advancement of biomedical imaging technologies, AI-driven diagnostic capabilities are poised to achieve qualitative leaps.

Established in 1993, the Asian Society for Cardiovascular and Thoracic Surgery is the largest professional academic organization in Asia specializing in cardiovascular and thoracic surgery. It ranks alongside the American Association for Thoracic Surgery and the European Association for Cardio-Thoracic Surgery, together constituting the three premier global academic forums in thoracic surgery. The invitation extended to Professor Zhang Lanjun’s team to attend the conference and deliver a plenary address signifies that the findings of their prospective study on an artificial intelligence diagnostic system have gained recognition from international peers. This achievement not only facilitates disciplinary advancement and international exchange but also accelerates the clinical implementation of novel artificial intelligence technologies.

AI in medicine is an entirely new interdisciplinary field combining medicine and engineering. Driven by technologies such as big data, artificial intelligence, and cloud computing, the growing capabilities of “AI medical assistants” have generated great anticipation within the medical community. The cross-sector integration of healthcare and technology is continuously advancing the practical application of AI in medical imaging, effectively linking AI, application scenarios, and value to ultimately serve clinical practice and benefit the public.