Lung cancer is the leading cause of cancer death worldwide that accounted for 18.4% of all cancer deaths.¹ In Hong Kong, lung cancer was associated with a crude mortality rate of 51.7% in 2018 and is considered the most common cause of cancer death.² Previously, population screening with low-dose computed tomography (CT) was shown to reduce the relative mortality of lunger cancer by 20%.³ Similarly, those who underwent volume CT screening were found to have a significantly lower lung cancer mortality rate than those who underwent no screening.¹ In the recent 2020 World Conference on Lung Cancer, Dr. Liu, Zai-yi of Guangdong Provincial People’s Hospital, Guangdong, China, explained the clinical challenges of adopting a large-scale lung cancer screening program and how artificial intelligence (AI) can enable effective and predictive analysis of lung cancers.

Traditionally, lung cancer is manually interpreted by radiologists through CT images to evaluate the nodule size, density and growth of the malignancy.⁴ As this assessment process requires heavy workload, a large number of radiologists would be needed to support large-scale screening programs. However, the rate of increase in medical image data had far exceeded the rate of increase in radiologists in both China and the United States since 2018, suggesting that such large-scale screening programs may be unfeasible with the traditional method. Clinically, misdiagnosis is also common where up to 36% of lung metastases cases were missed by the radiologists when evidence was present on CT scans.⁵ Together with considerable variations between radiologists, the subjective manual assessment approach has limited applicability in large-scale screening programs.

To improve the performance and inter-grader consistency, an AI platform can be trained to reproducibly and quantitatively identify the radiomic features of lung cancers. In 2016, the LUNA16 challenge has called for algorithms to detect lung nodules and found that the combination of classical detection algorithms and convolutional neural networks yielded the best detection results of lung nodules.⁶ Trained from the same LUNA16 challenge dataset, another deep learning algorithm that compared a patient’s current and prior CT volumes had achieved an area under the receiver operating characteristic of 94.4% for lung cancer risk prediction.⁴ Using these data-driven algorithms, an AI platform can help clinicians achieve efficient detection, precise quantitative analysis, intelligent qualitative analysis and convenient follow-up to lung cancer cases.

In the past few years, an increasing number of hospitals in China have adopted these AI platforms to improve diagnostic accuracy. Of the 100 Chinese hospitals interviewed, 98 had installed AI platforms to assist clinical diagnosis. In fact, 42% of these hospitals had installed 2 or more AI platforms to better support their day-to-day operation. Notably, there is also an increasing number of Chinese companies publishing China National Medical Products Administration and United States Food and Drug Administration approved AI platforms for lung nodules identification, which may contribute to the high adoption rate of AI platforms in China.

While these AI platforms can assist medical diagnosis, Dr. Liu noted that these AI platforms cannot replace radiologists when making critical clinical judgements. Sensitive to data variety, AI platforms may not produce a generalizable diagnosis due to the varying signal-noise ratio or reconstruction schemes of the imputed CT images. Economically, most AI platforms are in the start-up phase and are currently free-of-charge. Once these AI platforms are released in full version, the utilization of AI platforms may still be manpower efficient but not necessarily financially feasible. Ethically, the transparency and custody of the data collected by these AI platforms remain a debatable topic. The responsibility of misdiagnosis also remains controversial when the AI platforms indicate a different clinical interpretation than the radiologists. Finally, most AI platforms still rely on medical images only and would require the inclusion of other risk factors and clinical data to make more accurate diagnosis.

While it is easy to train AI systems, Dr. Liu remarked that some data such as family history and laboratory data are difficult to collect in sufficient quantity for AI training. Certain patient population such as the children may also have an insufficient sample size to generate an unbiased generalization. Qualities of medical images may vary between medical centers and require AI retraining before application. Standing on the perspective from both radiologist and AI user, Dr. Liu emphasized that data collection, labelling and management should be standardized, and doctors should work closely with software engineers to narrow the gaps in clinical judgement between AI systems and radiologists. “Most importantly, we need to be open-minded and apply our expertise with assistance from AI platforms to achieve precision healthcare for the lung cancer patients,” concluded Dr. Liu.