Thank you very much for the invitation. I have to thank the previous speaker for both setting the level and introducing the notion of AI and also for reminding us that reading a chest radiograph is a very humbling task. It is, truly. It's very difficult because it's projectional in multiple organs. So focusing now on this outline, I'm going to talk about biomarkers, imaging biomarkers, what the difference is between pre-specified and unspecified feature extraction, how to assess biomarkers if they're presented to you and maybe sold to you, and how to personalize lung cancer screening, hopefully, with that information. The biomarker definition, based on the best guidelines or statement issued by the FDA, clearly includes radiographic information. A very prestigious radiologist out of New York, formerly president of the RSNA, is famous for saying radiologists don't produce images, we produce data. And I'm highlighting this in this setting because I think for many years, biomarkers were deemed to represent blood-based biomarkers, clearly that is no longer the reality we live in. These are just the categories, and what we're looking for in the context of screening is that of either prognostic biomarker or diagnostic, even more so appropriate. Imaging biomarkers underpin cancer care across the globe. And what is really important is to understand that imaging, originally, and maybe hailing back to the chest radiographs we just witnessed, really were, imaging was meant to be a qualitative assessment. Things are up, things are down, we describe things with words. If you've ever met a radiologist, there are a ton of food analogies in everything we say. But that has shifted, and is continuing to shift to a quantitative analysis driven by a change in hardware and software. So our goal, collectively, in medicine is to move towards personalized medicine, and this is what we can achieve here. Now talking about pre-specified features. Pre-specified features are things that we can describe. So people refer to semantic features. A very commonly used feature in this context would be volume doubling time, as established here by the all-stars from the LCAP research team. But volume doubling time for pulmonary nodules, density of pulmonary nodules, borders of pulmonary nodules, all of those are things we can describe, and thereby those are known as semantic biomarkers. There are also things like agnostic biomarkers. Agnostic biomarkers are things that are hard for us to capture, but they're still pre-specified because some human expert, supposedly, came up with a formula to describe the relationship of pixel values and different densities and attenuation. So these are higher-order features, second- and third-grade features, also known as texture analysis. So these are the terms you'll see in the literature. Coming back to semantic biomarkers that you can identify on lung cancer screening chest CT, you'll have nodules, as already pointed out, but you also have body composition. Dr. Sandlin and her team did some very important work on that recently. Emphysema is a longstanding biomarker, but also interstitial lung disease. Sclerosis, things we can measure that have an impact on the likelihood of lung cancer because of their association with lung cancer risk. Most recently, pulmonary vasculature, as referenced down here. And all of these biomarkers can be extracted from CT images using quantitative image analysis, as simple as putting an electronic caliper on a nodule to measure the average diameter, but also things like computer vision, which lend themselves to those second- and third-degree extraction, kurtosis, and a lot of other features. Now, in contrast, there is the unspecified or unspecified features that we can extract from these CT images. The previous speaker introduced the notion of machine learning. So if you have supervised machine learning, meaning that you give the computer a set of scans, one with lung cancer and one without, and you supervise the computer by saying these are with, these are without, and then you let the computer figure out the formula of getting to the result, that would not then be forcing the computer to only identify and focus on features that you pre-select as the human, and it would enable the computer to maybe pick up features that are imperceptible to us as humans. Now, the downside of these convolutional neural networks, which is an architecture inside the deep learning space that lends itself to image analysis, is that they're often considered a black box. And black box because, as I said, we don't know how the computer is deriving the result, their prediction, either a classification task of saying lung cancer, yes, no, or even a continuous output such as future lung cancer risk, as is the case with Sibyl. Now, Sibyl, referenced here, leverages all the data due to this architecture, all the data contained on this Lotus chest CT volume, and there's no human input necessary. The no human input has tremendous implications for workflow. For all of these radiomics and pre-specified features, it is frequently necessary to identify the area of interest, identify the feature you want the computer to focus on. That is not the case if you lose, if you use this architecture or a machine learning approach. Now, interestingly, this is the data point here that's relevant to this discussion. If you were to take Sibyl and say, okay, we know that in year X there was a lung cancer, and let's say we know where that lung cancer is, and then you go back one year and there's no nodule, no perceptible nodule in that exact same location. And you force the model or you apply the model only to the scans without a precursor nodule visible, you still get a prediction of future lung cancer risk that is significantly better than chance. It drops compared to the all analysis where you give, where you allow the computer to allow to analyze all scans regardless if they're nodules or not. But if you limit it to that subset of no precursor nodule in the location of a future lung cancer, the predictive capability decreases, but it's still far better than chance, which suggests that the computer looks at features other than just nodules. Now, whenever you're presented with a biomarker, you need to ask about external validation. You need to ask about consistent performance on different patient populations, different scanners, image acquisition and reconstruction parameters. The image acquisition and reconstruction parameters are known to a specific or to a much larger degree affect radiomic approaches, texture analysis, less so the machine learning approaches. So that's what we did with civil recently. And then also ask the question of reproducibility. How can you in your own practice get those similar results? Is this a proprietary in-house software that someone wrote and locked away or is this a publicly available tool and publicly available either in exchange for money in a commercial platform or open access? So we strongly believe that open access is the way to go to increase the transparency, enable others to test our results. So we made the civil code freely available as on this link here. Now where does that lead to? Let's say you have a rigorously validated biomarker. You could potentially modulate low dose chest CT frequency and address some of the concerns about radiation that we heard about, some of the resource limitations we've heard about, some of the inabilities to really roll out screening because of cost concerns. After a baseline scan where risk is assessed, the subsequent screening intervals depending on risk as derived from the low dose chest CT imaging set could be modulated. Now the next steps, and this is my second to last slide, would be the necessary steps to get there. So one would be harmonization of image acquisition and reconstruction parameters. We know from large trials that that is possible. And we have seen a lot of efforts here to streamline, harmonize, put quality improvement parameters in place. So there is a way to get there. Pool annotated imaging data, for example, an effort by ISLAC here, the ELIC database as described by colleagues earlier, by Dr. Lam. And then I think the future is not only in CT derived biomarkers. One of the most successful models currently, the Brock model, uses both emphysema and nodule location as well as other epidemiologic factors, patient age and others we'll talk about in this session. But the idea that you have everything on the CT is probably incorrect. I think there probably will be an approach using some of the demographic factors, using the imaging information, and using blood-based biomarkers. All of this needs perspective evaluation, including a cost-effective analysis. Be mindful that with the civil analysis I showed you, there's no human input necessary when the model is deployed. No nodule measurement, no allocation of what lobe it's in. We can get that wrong sometimes. And then lastly, trustworthy AI, trustworthy AI in the sense of what are the privacy concerns, what are the concerns about drift of these algorithms over time. And lastly, and very importantly to all the radiologists in the audience, workflow integration. How do we get this model integrated into our workflow, which is very, very busy, and we heard about limitations of radiologists to workforce considerations. With that, I thank you very much.

AI in radiology

lung cancer screening

biomarkers detection

convolutional neural networks

medical transparency