“We perceive only a tiny portion of our environment, yet perception seems so very full. This illusion of fullness invites us to make wild cognitive leaps in blissful ignorance, and we repeatedly oblige” – Daniel Kahneman, 2002 Nobel Laureate.
In the next series of three blogs I’d like to share the motivation for prioritizing the development of Zebra’s first algorithms. We’ve generated five algorithms inspired by the potential benefit of “Lung Cancer Screening” CTs. None of our algorithms detect or characterize lung cancer, so we’ve been asked – fairly – what we were thinking. Here’s part one of what we were thinking.
It starts with appreciating how radiologists perceive medical images. We identify things and interpret them within context. The context may come from something else we see on the same CT or from a diagnosis in the patient’s medical record. The prevailing context in a given CT study is the specific reason for which the scan was ordered in the first place. Good radiologists will provide a thorough and accurate answer to the clinical question asked. Great ones will answer the question without being confined by it – they are able to assign importance to radiologic findings which no one would have thought to look for. These are the radiologists who see far beyond the ground below the proverbial “streetlight.”
But when do you stop? How many questions can be asked and answered in a given radiologic evaluation? We simply cannot ask all the right questions all of the time; we are quite happy to find one, two or three significant findings and then move on to the next study. This sentiment, which we radiologists call “satisfaction of search,” is what Daniel Kahneman describes as “blissful ignorance.” In short is is how we believe the “illusion of fullness” through which people experience the world.
We believe the first practical impact of machine learning in radiology will be in taking aim at the streetlight effect. Here are our first few examples.
Last year, Medicare began offering reimbursement for yearly chest CT scans among smokers over the age of 55 . A few years earlier, the American National Lung Cancer Screening Trial (NLCST) showed that CT screening would increase early detection of lung cancer and could decrease lung cancer mortality by 20% . Accordingly, to qualify for reimbursement, CMS stipulated that screening reports must use a standard terminology to indicate if any lung nodule is identified and the degree to which it appears malignant.
Few things capture the imagination better than saving a life by finding and resecting a small lung cancer before it metastasizes. Radiologists are asked to find the lung cancer, but is that the most important question? Is it the only question? Smoking is a risk factor for lung cancer, but it is also one of the strongest risk factors for cardiovascular disease – during the 8 year NLSCT study, cardiovascular disease accounted for more deaths than did lung cancer . Could the same Chest CT be used to screen for cardiovascular disease?
Similar screening trials in Italy , Israel  and the Netherlands  have all been used to measure coronary artery calcium as a marker of cardiovascular disease. The results suggest that routine coronary calcium measurements in lung CT screening would discover undiagnosed cardiovascular disease and trigger preventative therapy in 84 out of every 1000 (8.4%) participants. By comparison, undiagnosed lung cancer was detected in 3.9% of those who underwent CT screening in the NLCST.
And while coronary calcium is the most powerful known predictor of cardiovascular disease, it is not the only one visible on chest CT. The diagnosis of fatty liver (which can be made with nearly 100% accuracy on CT) is independently associated with a 2x – 4.6x risk of heart attack within seven years . Combining these two independent CT measurements – coronary calcium and liver fat content – could identify those in highest need of surveillance and preventative treatment.
This is why nearly two years ago, we recruited the first team of machine learning scientists to Zebra’s database and research platform with the goal of generating what we believe are five of the most practical and implementable medical imaging algorithms ever created. One provides an accurate and fully automatic assessment of coronary artery calcium on any chest CT. Another uses the same CT to detect fatty liver.
Two algorithms: two questions worth asking on every CT.
To be continued…
Eldad Elnekave, MD
Chief Medical Officer, Zebra Medical Vision
1. CMS. CMS LDCT Lung Cancer Screening Memo (CAG-00439N). https://www.cms.gov/medicare-coverage-database/details/nca-decision-memo.aspx?NCAId=274&NcaName=Screening+for+Lung+Cancer+with+Low+Dose+Computed+Tomography+(LDCT)&TimeFrame=7&DocType=All&bc=AQAAIAAAAgAAAA%3d%3d&. Published 2015.
2. Aberle DR, Adams AM, Berg CD, et al. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med. 2011;365(5):395-409. doi:10.1056/NEJMoa1102873.
3. Sverzellati N, Cademartiri F, Bravi F, et al. Relationship and prognostic value of modified coronary artery calcium score, FEV1, and emphysema in lung cancer screening population: the MILD trial. Radiology. 2012;262(2):460-467. doi:10.1148/radiol.11110364.
4. Shemesh J, Henschke CI, Shaham D, et al. Ordinal scoring of coronary artery calcifications on low-dose CT scans of the chest is predictive of death from cardiovascular disease. Radiology. 2010;257(2):541-548. doi:10.1148/radiol.10100383.
5. Jacobs PC, Gondrie MJA, van der Graaf Y, et al. Coronary artery calcium can predict all-cause mortality and cardiovascular events on low-dose CT screening for lung cancer. AJR Am J Roentgenol. 2012;198(3):505-511. doi:10.2214/AJR.10.5577.
6. Osawa K, Miyoshi T, Yamauchi K, et al. Nonalcoholic Hepatic Steatosis Is a Strong Predictor of High-Risk Coronary-Artery Plaques as Determined by Multidetector CT. PLoS One. 2015;10(6):e0131138. doi:10.1371/journal.pone.0131138.
7. Puchner SB, Lu MT, Mayrhofer T, et al. High-Risk Coronary Plaque at Coronary CT Angiography Is Associated with Nonalcoholic Fatty Liver Disease, Independent of Coronary Plaque and Stenosis Burden: Results from the ROMICAT II Trial. Radiology. 2014;274(3):140933. doi:10.1148/radiol.14140933.
8. Pickhardt PJ, Hahn L, Muñoz del Rio A, Park SH, Reeder SB, Said A. Natural history of hepatic steatosis: observed outcomes for subsequent liver and cardiovascular complications. AJR Am J Roentgenol. 2014;202(4):752-758. doi:10.2214/AJR.13.11367.