I would like to thank the organizing committee for inviting me to speak on AI with chest x-ray. I have no disclosures. As we have heard throughout the whole day, lung cancer is a significant health burden. It causes more cancer deaths than prostate, breast, and colon combined. And for prostate, breast, and colon, we have screening tools. However, in Singapore, lung cancer screening is still not a national program. And without a significant stage shift, it is unlikely that the five-year survival in Singapore is going to improve because majority of them, when they present with symptoms, 65% of our lung cancer patients are in stage four. So we have seen the cancer statistics globally in 2018. Lung cancer constitutes the second most common in men and females. However, it is the top killer for both males and females. And in Asia, Asia has the highest lung cancer burden in the world, 60% incidence and 62% mortality. And if we look at the smoking prevalence of lung cancer patients by gender in Asia versus the West, we can see that there is a high prevalence of non-smoking lung cancer in Asia, accounting for more than 30%. And less female lung cancer is related to smoking in Asia. In Singapore, lung cancer is the third most common cancer in men and women. However, it is the leading cause of cancer death in men and the third highest in females. And if we look at the low and middle income countries, other than the red, it represents pretty much most of the world except the red areas. If we look at the patient exposure from radiologic and nuclear medicine procedures worldwide, chest x-ray is the most commonly requested radiological study. And both the high and low income countries order chest x-rays very commonly. And that is because it is widely available, is associated with very low radiation dose and is cheap. The cons of a CT scan is that it is associated with a higher dose of radiation, is not widely accessible and low income countries or middle income countries cannot afford CT scans. And with increasing use of CT, it also exposes a higher proportion of false positives and incidental findings that will lead to additional lab testing, increasing the patient anxiety as well as harm associated with downstream investigations. So if we look at the world, CT examinations are more commonly performed in high income group countries as compared to low and middle income group. But interestingly, if you look at America, where a CT examination constitutes 18% of the global consumption, the average individual radiation dose exposed per patient is about 2.2. So there is a little bit of concern with regards to the amount of radiation that's associated with a CT screening. So if we follow the UPSTF criteria and if a 55-year-old participant undergoes a low dose CT screening over a period of 20 to 30 years, the cumulative radiation exposure would range between 280 millisieverts to 420 millisieverts. And this exposure far exceeds those of nuclear workers and atomic bomb survivors. So it is likely that with the current lung cancer screening protocols, if conducted over 20 to 30 years, can independently increase the risk of lung cancer beyond cigarette smoking as a result of cumulative radiation exposure. However, from the PLCO trial, it has conclusively proven to us that annual screen with chest x-ray does not decrease the lung cancer mortality when compared to usual care. So why is it so that it's so challenging to look for cancers on chest x-ray? In this study that looked at about six years, no, eight years, where they found 40 cases of non-small cell lung cancer that were evident on chest x-ray retrospectively but undetected by radiologists, they found that majority of them measured about 1.9 centimeters, often located in the upper lobes, with the right upper lobe more commonly missed than the left upper lobe. And if we look at the segments, the apical posterior segments tend to be more often affected. If the clavicle overlies the nodule, it can obscure it, as well as small nodules in the peripheral locations where the ribs can also obscure it. So hence, there is an article that is really very interesting that tells us when, why, and where cancers can be missed on chest x-ray. The first is the observer factor, where the observer error is the most significant, and it comprises of scanning error, recognition error, decision-making error, and satisfaction of search. If we do it in a haphazard manner, or the radiologist scans the x-ray in a haphazard manner, it is no wonder that nodules may be missed. However, if you scan in a regular manner, as shown here, recognizing that there are blind zones, which are usually in the apices, around the hilum, and sub-diaphragmatic, by having this awareness, we can formulate strategies to reduce observer error and methods to improve the technique, and perhaps using AI as an automated detection may be valuable in decreasing the likelihood of missing the lung cancer. So, for example, looking at this CT scan, the radiologist missed a right lower zone nodule, interpreting it as a nipple shadow, and detected this left upper lobe small nodule, and because of the satisfaction of search, meaning that he or she has identified abnormality in the left upper lobe, a CT scan was performed, which showed two lesions, both in the left upper lobe, as well in the right superior segment of the right lower lobe. However, the left upper lobe nodule, six months later, did not increase in size. In fact, the right lower lobe, which was missed on reading of the chest x-ray, grew in size, and when this was biopsied, it was adenocarcinoma. There are some tumor characteristics that can affect missing the nodule on chest x-ray, particularly the size. Size matters, and if the nodule measures between 10 to 30 millimeters, it gives us an error of about 29 percent, however, if the lesions are more than four centimeters, these nodules are often not missed. Conspicuity of the nodule, meaning the contrast of the margins against the surrounding structure, that is also an important factor for us to see the nodule, and also distracting issues. So for example, when we look at this x-ray, there is a left pneumothorax as a result of the pacemaker implantation, and hence the radiologist did not report a right upper zone nodule that was very close to the hilum, to the medial spinal area, and on the lateral chest x-ray, because it was obscured by the thoracic vertebra, this was also missed. The patient subsequently developed pulmonary embolism, and a CT scan was actually done for it, and this showed that there was a soft tissue mass in the right upper zone. So if the nodules are around the hilum, this may be also a source of error because of surrounding overlying anatomical structures. So in fact, Mum et al found that 65 percent of pulmonary lesions originating in the hilum were overlooked in a screening program, and centrally located tumors tend to be larger than peripheral ones. So cardiac structures may also obscure the left lower or retrocardiac nodule or mass, and this can be overcome with the lateral chest x-ray, where an increase in attenuation, also known as the positive posterior sign, has a good predictive value. And if we do a chest x-ray in a supine position with an AP projection, as shown here, it is no wonder that apices become a blind spot, and indeed there was actually a nodule in the right upper apex, but because of the technique of the projection, this was missed, and on CT scan it confirms the presence of the nodule. So now I'm going to talk about artificial intelligence, whether the role of artificial intelligence combined with chest x-ray could circumvent all these errors that we encounter in our x-ray reading. So artificial intelligence is really an umbrella term for techniques that simulate human intelligence through computational systems. It is divided into machine learning, where there are algorithms that build computer models capable of learning and making predictions or decisions from data, and deep learning is part of a machine learning where it uses deep neural networks to automatically learn from data without human intervention. So this study actually looks at a commercially available AI algorithm, also known as Red Dot, which has convolutional neural networks. So what it did was to have this AI as a triage, as a first reader. So this shows that it performs very well, and when combined with the radiologist, it actually increases the sensitivity, specificity, and accuracy in determining the abnormal lesions. So for example, this is the AI view of the chest x-ray where there is an obvious pleural effusion, but two out of three radiologists did not notice a right upper zone nodule that is identified on the AI. So I'm just going to end because it's almost time. So they have looked at AI complementing radiologists, and it shows that it levels the playing field, meaning that resident radiologists, residency radiologists also can do as well as senior radiologists with the help of AI, as shown here on the heat map. And then the last is using AI to create a prediction score, as shown here, where they can get the EMR details like age, sex, as well as smoking, and a chest x-ray without specific smoking pachyars. And they have developed a chest x-ray LC model, which performs as well as the PLCO model in identifying patients who are eligible for CT screening. So with that, thank you very much for your attention.

lung cancer

early detection

artificial intelligence

screening challenges

CT scan limitations