Since its introduction in 1973, X-ray computed tomography (CT) has become a leading modality for diagnostic imaging. The advantages of CT are manifold. Above all, they include rapid scanning and small spatial resolution, which allows for relatively quick and accurate diagnosis of injuries and disease. CT has also been an imaging tool of choice for the staging and treatment follow-up of cancer.

Growth in use, but variations between countries
Overall, CT use has grown rapidly. The total number of scans in the US is estimated to be in the region of 80 million a year. In England, the National Health Service (NHS) reported 4.8 million CT scans in 2016/17, which is 40 percent more than the 3.4 million MRI scans done during that year. CT usage has also been growing rapidly – in England at about 8% annually, compared to just 1.5% for X-rays and 5% for ultrasound.
Nevertheless, there are significant variations between countries in the intensity of CT use. According to data from the Paris-based Organization for Economic Cooperation and Development (OECD), the annual rate of CT scans per 1,000 inhabitants ranges from a high of 225-230 in the US and Japan, to a low of 37 in Finland. The rate is about 80 in Italy, 90 in the Netherlands, 110 in Spain, 140 in Germany and 200 in Belgium and France.
Differences in radiation dosing practice
Though large, such divergences are considered to be less significant than differences in radiation dose to which patients are exposed, for the same condition. In December 2007, a study published in ‘European Radiology Supplement’ had found dosage could have been halved in many cases without impacting on image quality. Another study two years later revealed a 13-fold difference between the lowest and highest radiation doses used for identical CT procedures by four clinical sites in the neighbourhood of San Francisco.
Concerns about such issues have been dramatically highlighted after a major new international study, which attributes differences in dosage to the person doing the scanning rather than to patients or equipment. The study, published in ‘The British Medical Journal’ (BMJ) in January 2019, found that patient characteristics, make and model of scanner, and type of hospital where the CT scan was done had little effect on the amount of radiation used.

Analysis of 2 million CT scans in 151 institutions
The BMJ study was based on a massive effort by a research team led by Dr. Rebecca Smith-Bindman, a professor in the Department of Radiology and Biomedical Imaging at the University of California San Francisco (UCSF). The researchers analysed dose data for over 2 million CT scans of the abdomen, chest and head, at 151 institutions in seven countries.
Their findings are likely to resonate strongly, given the association of radiation with cancer. Although CT scans account for a minority of diagnostic radiologic procedures, they use large amounts of radiation per image. Some estimates suggest that CT contributes nearly half the US population’s radiation dose from all medical examinations. The figure in England is higher, at 68 percent, although plain radiography is used five times more often than CT in the country (22.9 million procedures in 2016/17 versus 4.8 million).

Cancer risks of CT
The association with cancer has been controversial, especially when predictions of the impact of CT scanning have been based on a linear-no-threshold dose-response model. Some have argued that CT radiation doses are too low to produce any health effect.
There is also uncertainty about how to calculate risk accurately. This is because of a host of factors. Firstly, radiologists are not necessarily familiar with CT radiation exposure descriptors (volume CT dose index and dose length product). Secondly, there have been a series of revisions about the relative sensitivity of organs to radiation. Finally, radiation dose in units such as millisieverts (mSv) are used to estimate population risks based on generic models, not individual patient calculated dose. Indeed, the radiation dose in a typical CT scan (1–14 mSv depending on the exam) is similar to the annual dose received from natural sources, such as radon and cosmic radiation – which typically varies from 1 to 10 mSv, depending on where a person lives.

Even small risks justify search for solutions
Nevertheless, the current consensus is that, even if the risk of cancer from CT imaging is small, the economic burden of treatment of the proportionately reduced number may well be significant, given the high prices of cancer treatment.
Neither does anyone question the logic of attacking even a small cancer risk. In December 2009, a report in ‘The Archives of Internal Medicine’ made a detailed assessment of projected cancer risks due to CT scans in the US. The study was conducted by a team from the Radiation Epidemiology Branch of the National Cancer Institute (NCI), and argued that changes in practice might help to avoid the possibility of reaching an attributable risk of 29,000 cancer cases based on CT scans in the year 2007. The authors also observed that the impact would be largest in abdomen, pelvis and chest CT scans in adults aged 35 to 54 years.

Unnecessary scans
One of the most vexatious issues concerns CT scans which are not medically necessary, especially when it concerns repeat imaging of a particular patient – and the ensuing enhancement of cancer risk. According to one estimate, unnecessary scans could account for as much as 30 percent of CTs in the US. In Europe, such a figure is also likely to be high in countries such as Belgium and France where per capita CT scan levels are close to those of the US.
Though the US state of California has passed a law requiring documentation in a patient’s medical record of radiation dose used for every CT scan, compliance has been inconsistent. Perspectives in Europe are problematic too. For example, the European Union collects dose levels in Europe, but there are major differences in definitions and data collection techniques.

Progress in pediatric dosing
Until the NCI study at the end of 2009, the emphasis on reducing CT cancer risks had largely been on pediatric scans. The authors of that paper noted there was evidence of pediatric doses being reduced as a result of social marketing campaigns such as Image Gently. The latter was launched in 2008 by the Alliance for Radiation Safety in Pediatric Imaging.

Lessons from the pediatric dose control campaign
One of the key recommendations of Image Gently was to promote standardization of pediatric dose measurements and display across vendor equipment.
This is precisely what the recent BMJ study proposes to do for all patients. The authors of the study assessed mean effective doses and proportions of high dose examinations (defined as CT scans with doses above the 75th percentile defined during a baseline period) for abdomen, chest, combined chest and abdomen, and head CT. These were classified by patient characteristics (sex, age, and size), type of institution (trauma centre, 24×7 care provision, academic and private hospital), practice volumes, machine manufacturer and model, country etc. The figures were adjusted for patient characteristics, using hierarchical linear and logistic regression.
For example, after taking into account patient factors, a fourfold range in radiation doses still existed in abdominal scans. Similar variations were found for chest and combined chest-and-abdomen scans.

Huge variations in dose
The BMJ study found that variations in radiation dose across institutions and countries were huge. For abdomen CT examinations, the mean effective radiation dose differed by a factor of four, with a 17-fold range in the share of high dose examinations (4 to 69%). Variations in mean effective dose for chest scans and combined chest plus abdomen scans were also close to four times, while the share of high dose exams varied from 1 to 26%, and 2 to 78%, respectively. For head CT, the differences were less spectacular (with the range of mean effective doses less than 1.5 times and the share of high dose exams ranging from 8 to 27%.

Achievable and universal standards
However, when the UCSF group adjusted for technical parameters, that is, in terms of the way CT scanners were used by medical staff, the variations in doses nearly disappeared.
The researchers conclude that it is possible to optimize doses to a “single set of achievable quality standards” and apply this “to all hospitals and imaging facilities.” They also noted that the choice of “appropriate CT protocol parameters might be less complex than widely believed.” The key to protocol optimization lies in updating physician awareness and recalibrating expectations about what constitutes a diagnostic CT scan. The latter will be based on a better alignment of CT protocol parameter choices with diagnostic image quality requirements.
One interesting finding was that institutions with lower average doses shared scanning approaches. These institutions tended to limit the number of protocols, with each relying on the minimum dose required to answer the clinical question. They used multiple CT scanning infrequently, had lower settings for tube current and tube potential, and used higher pitch for most, if not all, imaging indications.

The way ahead
The road to CT dose reduction and standardization will vary by type of institution and country. This is due to differences in the make and model of CT scanners as well as medical cultures, in terms of radiologist preferences and personnel support. There are case studies of protocol overhauls taking a year or more, and needing to be kept up-to-date with new CT software and scanner upgrades. Examinations with higher radiation exposure generally give more acceptable images than those where exposure is lower. The challenge is to optimize a ‘correct’ minimum dose for different patient sizes, ages and conditions. Continuing improvements in scanning technology will undoubtedly also be part of the process of optimizing protocols. On their side, some companies have been experimenting with artificial intelligence algorithms to position patients correctly in a CT scanner. Off-centre CT scans can expose patients to much higher levels of radiation
than necessary.

40th 

In February last year, the US Food and Drug Administration (FDA) cleared the first medical device which uses artificial intelligence (AI) to provide clinical decision support for stroke. The Viz.AI Contact application uses an AI algorithm to identify a suspected stroke and notifies a specialist more quickly than was previously possible. Faster treatment, in turn, lessens the extent of a stroke or its progression. Subsequent FDA clearances and a recent decision to formalize regulations for such evaluations are likely to stimulate further innovation and acceptance of AI devices.

Saving time
Viz.AI Contact analyses CT images of the brain and sends a text notification by smartphone or tablet to a vascular neurologist or a neuro-interventional specialist, should a large vessel occlusion (LVO) be suspected. The algorithm automatically notifies the specialist at the same time that a review of the images is being conducted by a first-line provider. This is faster than the usual standard of care where patients wait for a radiologist to firstly review CT images and then notify a neurovascular specialist.

Retrospective study and real world data
Viz.AI, Inc., which developed the Contact application, submitted a retrospective study of 300 CT scans. This compared the performance of the image analysis algorithm and notification functionality against two trained neuro-radiologists.
Real-world evidence from a clinical study demonstrated quicker notification of a neurovascular specialist, in cases where blockage of a large vessel in the brain was suspected. In more than 95 percent of cases, the automatic notification was faster, saving an average of 52 minutes (with a range of between 6 and 206 minutes).

De Novo premarket review
The Viz.AI application was reviewed by the FDA through its De Novo premarket review process, a regulatory pathway for new types of medical devices that carry low to moderate risk, but lack a legally marketed predicate device to base a determination of equivalence. The FDA action creates a new regulatory classification, allowing other devices with the same medical imaging intended to obtain marketing authorization by 510(k) notification. One of the first areas to benefit from Viz.Ai will be AI or computer-aided triage devices, whose potential in fields such as emergency medicine is likely to be vast. Viz.AI, Inc., itself is developing Viz ICH, which uses AI to automatically detect intra-cerebral hemorrhages and triage the patient directly to the neurosurgeon on call.

Decision support for breast cancer screening
Nine months after FDA approval of Viz.AI, at the 2018 Radiological Society of North America (RSNA) annual meeting in November, Siemens Healthineers showcased the AI-based features of syngo.Breast Care, a mammography solution. syngo.Breast Care aims to provide interactive decision support for breast cancer screening.
Transpira, Siemens’ mammography reading software, is based on deep learning techniques, with training provided via over 1 million images. As a result, syngo.Breast Care’s AI-based algorithms evaluate and interpret individual lesions as well as 2-D mammograms and 3-D tomosynthesis. The system also sorts and scores cases on a 10-point scale, based on radiologist preferences of risk factors such as lesions, micro-calcifications and other abnormalities.
Siemens Healthineers aims to integrate interactive decision support into syngo.Breast Care, and reduce radiologists’ workload for the interpretation of mammograms. This has become especially challenging, given rapid growth in the use of techniques such as 3-D breast tomosynthesis.

Small firms also in play
Smaller firms have also targeted this area. ICAD’s ProFound AI, for example, also leverages AI to detect cancer in breast tomosynthesis. The software, which was FDA cleared less than a month after syngo.Breast Care was unveiled, examines every image in a tomosynthesis scan, detects malignant soft tissue densities and calcifications.
Profound AI estimates a ‘Certainty of Finding’ for each detection and, like the classification system in syngo.Breast Care, assigns Case Scores to each case to represent confidence that a detection or case is malignant. The scores are represented on a scale from 0 to 100 percent, with higher scores indicate high confidence levels in malignancy. This, in turn, is expected to improve detection, lead to fewer patient recalls and save mammographers time in reading images. This makes it geared toward screening, although it can evidently be used for diagnostic studies.

AI at inflection point
The above examples demonstrate that the use of AI is now close to an inflection point in terms of clinical decision support tools. These will provide physicians usable interactive and dynamic pathways which move beyond decision support to true evidence-based decision making, along with personalized care recommendations.
To many experts, AI seems to have been the missing link for tools that assist radiologists in improving appropriateness of follow-up recommendations for incidental findings, and thereby to enhance adherence to guidelines available at point of care. One of the consequences of such AI-assisted tools will be to reduce the variability in follow-up recommendations, as well as unnecessary imaging studies.

Diagnosis and decision support versus analysis and detection
Maximum attention to AI in imaging is currently on diagnosis and decision support. AI in areas such as quantitative analysis and assisted detection can be considered a spin-off from automation, which has been around for a longer period of time, but reinforced more recently by machine learning.
Automated quantification tools are now sufficiently mature and routinely accepted in the market. AI algorithms are used to make measurements from imaging exams and perform calculations which were previously manual and time-consuming. AI-driven quantitative analysis tools also are being used in data analytics for data mining electronic medical records, billing systems, patient scheduling and even in stand-alone scanners. Mined data range from radiation dose used by particular technologists for specific protocols to predictive analytics that pinpoint spikes in demand by day and time, and schedule back-up staff in the radiology department.
By contrast, the application of AI (and even automation) in medical fields such as computer-aided diagnosis and clinical decision support is very recent, and is likely to be some time before they become commonplace. The principal focus on AI use for image diagnosis is where timing is crucial – such as a heart attack or stroke (e.g. Viz.AI Contact). Closely related areas include tools to reduce review time for complex exams, and help triage patients needing more immediate care or other kinds of back-up.

Other new AI imaging applications
One exciting new entrant into AI in imaging is IcoMetrix, from Belgium’s IcoBrain. This FDA-cleared algorithm analyses CT scans to characterize traumatic brain injury, using deep learning to quantify the severity of such typically qualitative indicators of brain injury as hyperdense volumes, compression of the basal cisterns and midline brain shift.
Another FDA-cleared device is Cardio AIMR, which analyses MR images for cardiovascular blood flow. Its developer, Arterys, also has other AI tools to measure and track liver lesions and lung nodules, accelerate display of medical images, and interface with the common desktop Google Chrome browser to display mammograms.

The challenge of integration
Although the FDA is clearing the way for follow-on AI products, there are concerns that the process is constrained to highly specific medical imaging diagnostic reviews. Some radiologists are questioning the viability of new AI software systems, if they require scores of different contracts and integration into a hospital or enterprise imaging system – which would be a problem not only for hospital IT departments but also for legal review.
One of the ways forward is by reconfiguring approaches to enterprise imaging by streamlining workflow. Some vendors are developing bridges between different AI applications. One of the immediate goals is to have AI imaging dovetail into picture archive and communication systems (PACS) as well as vendor neutral archives. For example, Viz.ai software is designed to receive DICOM images directly from any CT scanner to a local virtual machine (VM) behind a network’s firewall.

Major firms nurture start-ups
Leading healthcare technology vendors are also starting to actively partner with smaller companies to provide a combination of in-house and third-party apps via a web-based AI app store platform. One good example of this is Siemens’ Digital Ecosystem, which offers an online menu of apps from Siemens and its partner, including some offering AI-enabled technology. Similar AI app store initiatives are also being taken by other vendors.
At RSNA 2018, where Siemens showcased syngo.Breast Care, IBM Watson said it would begin to partner with AI vendors to offer products on its new AI Marketplace, by offering standardized application programming interfaces (API) for building or integrating third party software and making it available through the IBM Cloud. Smaller vendors have seized such opportunities. French imaging agent vendor Guerbet, for instance, is working with IBM Watson Health to develop AI software to support liver cancer diagnosis and care.
IBM had initially planned to develop and launch its own AI solutions across the healthcare spectrum. However, it had to cope not only with delays in commercializing its own AI products, but small and nimbler start-ups, such as viz.AI getting ahead in obtaining FDA clearance. The biggest setback was MD Anderson ending its partnership on cancer imaging with IBM.
Other major players are also treading similar paths. GE Healthcare’s Edison platform is designed to help accelerate the development and adoption of AI and other new technologies, with clinical partners using Edison to develop and test algorithms and mate them to Edison applications and smart devices. On its part, at RSNA 2018, Philips Healthcare also launched its IntelliSpace Discovery 3.0 visualization and analysis platform to prepare patient data to train and validate deep learning algorithms. The platform is designed specifically to support imaging research.

FDA to formalize De Novo rules
Developments in AI-enabled clinical decision support, like broader AI healthcare applications, are likely to pick up after the FDA decided to formally establish regulations for the De Novo classification process in December 2018. Although the De Novo process is part of the Food and Drug Administration Modernization Act, the FDA Safety Innovation Act and the 21st Century Cures Act, it is currently not covered by any specific regulations. If finalized, the proposed rules are intended to provide clarity and transparency on the De Novo classification process.