Lung Nodule AI at a Crossroads: Where to Next?

“China is running out of lung nodules.”

Upon learning that the PACS (Picture Archiving and Communication Systems) of a renowned Grade 3A hospital in Southwest China had integrated lung nodule AI imaging diagnostic systems from seven artificial intelligence companies, a radiologist jokingly commented on the current situation.

Lu Xun once said, “There was no path in the world; it became a path only after many people had walked it.” However, in the field of AI-based detection of pulmonary nodules, Chinese startups have already paved a well-trodden highway akin to the Beijing–Shanghai Expressway. In core hospitals where numerous AI vendors compete fiercely, it has become commonplace for a single radiology department to host servers from four to five different AI companies.

As of July 2018, incomplete statistics show that more than 20 AI companies have launched specific products in the field of pulmonary nodule screening alone. Most of these companies have secured venture capital funding, while a growing number of overseas-educated PhDs, research institutes, and hospital expert teams are gearing up and eager to enter the market.

Everyone is heading toward the same goal.

From the demand side, China ranks first globally in both the annual number of newly diagnosed lung cancer cases and the annual number of lung cancer-related deaths. This has created a strong demand for early screening, leading to the widespread promotion of low-dose spiral CT. In terms of image quality, chest CT images feature thin slices, clear fields of view, minimal interference, and recognizable lesion characteristics, making them ideal for intelligent image interpretation. Coupled with the scarcity of radiologists in China and strong support from national policies, the foundational conditions for application in this field are virtually perfect.

However, the demand in any disease area in China appears immense due to its large population base. What truly propelled the rise of intelligent detection of pulmonary nodules was the advent of the big data era.

In July 2017, Pranav Rajpurkar, the lead of SQuAD, publicly released his algorithm for detecting pneumonia from chest X-rays and unveiled the largest publicly available chest radiography dataset at the time, which contains over 100,000 frontal-view X-ray images covering 14 different diseases.

In October 2017, the NIH (National Institutes of Health) Clinical Center released a dataset comprising over 30,000 patients and more than 112,000 frontal-view X-ray images, along with image labels for 14 disease categories extracted from radiology reports using NLP technology, making it freely available to researchers worldwide.

In July 2018, the NIH (National Institutes of Health) Clinical Center once again shared a large-scale CT image database, comprising 32,000 CT scans and associated imaging data demonstrating various diseases. This initiative aims to help scientists and clinicians enhance their skills in radiological diagnosis of diseases, while also allowing AI developers to freely utilize these datasets for training AI systems.

For entrepreneurs at Chinese medical AI companies, these public U.S. datasets serve as an ideal incubator, providing abundant data “fuel.” The prevailing approach among China’s medical AI firms is to first train AI models to a certain level of accuracy using public datasets, then deploy them in clinical settings and further enhance their performance by “feeding” them with China’s vast troves of chest CT images.

“For entrepreneurs, the most difficult resource to obtain is clinical medical data. No matter how robust their technological foundation or how many algorithmic models are available for reference, without clinical data to train AI, all efforts are like water without a source.” Thus stated an operator of an imaging cloud platform. “China’s vast base of CT data has enabled its artificial intelligence technology to advance at a pace no less impressive than that of its birthplace—the United States.” A 2018 article in The New York Times (NYT) noted that China’s Alibaba and Yitu had already introduced AI into the healthcare industry ahead of Amazon.

But at the same time, he cautioned that although AI entrepreneurs in China start on similar tracks, this does not mean they will compete on the same dimension in the long run.

“Just as the achievements of college graduates can vary dramatically ten years after graduation, so too do factors such as R&D capability, depth of understanding of the healthcare industry, strategic foresight, and scale of industrial layout significantly influence the development prospects of AI companies once they choose intelligent pulmonary nodule detection as their breakthrough point,” this executive remarked.

In 2017, major AI companies specializing in pulmonary nodule detection delivered impressive results, with sensitivity rates soaring to 95%, 96.5%, 98.8%, and beyond. Subtle pixel-level differences imperceptible to the human eye were rendered unmistakable by the powerful computational capabilities of AI.

However, medicine is not a field where problems can be solved merely by mastering basic functions.

Currently, to ensure the accuracy of image interpretation in clinical practice, it is standard for a licensed physician and an associate chief physician to jointly review the same patient’s chest X-ray. After the radiologist completes the initial reading, a senior physician must re-examine the images and sign off to confirm the findings. The purpose of AI is precisely to replace the first step in this process. AI systems not only possess exceptional “visual acuity,” enabling them to detect nearly every tiny nodule, but they also do not suffer from fatigue or visual strain, capable of processing thousands of chest CT scans in mere milliseconds.

By continuously enhancing the sensitivity of AI, it is theoretically possible for AI to detect every pulmonary nodule. However, the consequent high false-positive rate poses a significant challenge. Should we rapidly maximize sensitivity while temporarily disregarding the false-positive rate? Or should we invest greater efforts to simultaneously elevate both sensitivity and specificity to clinically usable levels? Alternatively, are there more scientific and accurate evaluation metrics available?

Professor Gong Xiangyang, a renowned radiologist in China and Director of the Department of Radiology at Zhejiang Provincial People's Hospital, once stated,Balancing specificity and sensitivity is challenging; therefore, many companies prioritize sensitivity during system development, aiming to enhance specificity while ensuring high sensitivity.

“Achieving both high sensitivity and a low false-positive rate is indeed challenging, rigorously testing the technical prowess of AI companies. It requires maintaining sufficient sensitivity while ensuring a sufficiently low false-positive rate, so that the vast majority of detected nodules are correct and clinically significant. This task is far more difficult than it may appear externally. Yitu Healthcare has made arduous efforts to achieve industry-leading performance in both aspects,” said Ni Hao, President of Yitu Healthcare, in an interview with the media. “An excessively high false-positive rate would significantly increase physicians’ workload and undermine the original intent of using AI to support clinical practice.”

Ni Hao also cautioned that prioritizing sensitivity alone while neglecting the false-positive rate would not only increase physicians’ workload in verifying imaging reports but also impose significant psychological stress on patients, potentially leading to overtreatment driven by panic, thereby wasting medical resources and exacerbating patient burdens.

To more objectively reflect the facilitation of clinical work by AI products, Yitu Healthcare has introduced a new industry-wide metric—the clinical adoption rate of structured reports.

Specifically, this metric comprises two aspects: “Structured Reporting” and “Clinical Adoption Rate.” “Structured Reporting” requires pulmonary nodule AI not only to detect nodules but also to include information such as nodule size, morphological description, and features suggestive of benign or malignant nature, thereby generating a structured clinical report. The “Clinical Adoption Rate” is a more stringent criterion—what proportion of the structured reports generated by the AI can be directly adopted by clinicians without any modification?

Yitu Healthcare announced their care.ai^TMClinical Feedback on the Intelligent Imaging Diagnostic System for Lung Cancer in the First Half of 2018: The Clinical Performance Metric Reached 92%. This is an exceptionally outstanding result, indicating that 92% of the structured reports generated by the AI system were approved and directly adopted by radiologists. Behind this figure lies the immeasurable amount of working time saved for radiologists by the AI system.

“Only by embedding artificial intelligence into physicians’ clinical workflows—particularly when clinicians trust our diagnostic reports—can AI meaningfully enhance clinical efficiency. While sensitivity is a critical metric, it is merely the starting point,” said Zheng Yongsheng, Vice President of Products at Yitu Healthcare. “Currently, leveraging this proprietary technology, the system has been adopted by more than 100 Grade A tertiary hospitals across China and is being comprehensively deployed in primary-care hospitals, where AI-driven healthcare support is most needed.”

As the detection rate of pulmonary nodules gradually approaches its theoretical limit, all medical AI companies are pondering: where will intelligent diagnosis of chest CT scans head next?

At its core, this question asks how AI capabilities can advance from answering “what is seen” to addressing “what it is” and “how to treat it.”

“On one hand, we continue to delve deeper into pulmonary nodule detection, advancing from image interpretation to MDT decision-making. On the other hand, we are expanding from the sole detection of pulmonary nodules to the intelligent diagnosis of various lung diseases, such as pneumonia, tuberculosis, chronic obstructive pulmonary disease (COPD), and bronchiectasis. By addressing departmental workflow challenges through a single application, we aim to break free from the constraint of ‘one AI model per disease,’” said Ni Hao. “This path will be long and arduous, but only by continuously meeting clinical needs can we truly become valuable assistants to physicians and drive the development of future smart hospitals.”

On June 15, Yitu Healthcare, in collaboration with West China Hospital—one of China’s top-tier tertiary hospitals—unveiled the world’s first intelligent multidisciplinary diagnostic system for lung cancer. Hailed as the AI application “most akin to clinical reasoning,” this system goes beyond basic functions such as nodule screening to provide comprehensive diagnostic coverage for all types of lung cancer lesions. By integrating multidisciplinary clinical information, it delivers holistic diagnoses based on the latest domestic and international clinical guidelines for lung cancer diagnosis and treatment. As the volume of clinical cases increases, the system will continue to evolve in intelligence and sophistication, serving as a valuable tool for primary care physicians to enhance their diagnostic and therapeutic capabilities while reducing misdiagnosis and missed diagnoses.

More encouragingly, certain features of the product have already entered clinical trials, undergoing rigorous clinical evaluation. If successful, this would represent a significant breakthrough in medical AI.

“Just like all the technological revolutions experienced in human history, artificial intelligence will integrate into doctors’ workflows and collaborate with the medical community to better benefit patients,” said Professor Liu Shiyuan, Chairman of the China Medical Imaging AI Industry-Academia-Research-Application Innovation Alliance and Director of the Department of Diagnostic Radiology and Nuclear Medicine at Changzheng Hospital, Second Military Medical University. “The future is already here; it’s just unevenly distributed.”