The market for artificial intelligence (AI) based applications is heating up throughout the healthcare enterprise, with an ever-increasing number of offerings for use in diagnostic imaging. Numerous products have achieved FDA clearance and CE marking over the last years. Many, if not all, of the imaging centric approvals to date are for narrow indications that are best described as features rather than products from the perspective of a clinical imager. While the much-heralded ability to assess pediatric bone age from skeletal radiographs could be marketable as a free-standing product, the interested market is small. Imagen achieved approval for wrist fracture but won’t have a genuinely marketable product until it transfers learning and gains broader, full skeletal application approvals. Mammography represents a huge opportunity for AI with ever increasing study volumes/complexity and a shortage of qualified readers. Curemetrix has regulatory approval to promote positive cases in a stack of screening mammography studies but, on its own, the practical impact is small. They will not really have a complete offering until they gain approval for the full and functional CAD capabilities that published research suggests can improve cancer diagnosis and clinical workflow, due to fewer flagged false positives compared to conventional offerings.

Innumerable concerns are efforting the AI based identification of what computer scientists call anomalies on chest x-rays like pneumothorax, pneumonia and misplaced lines andtubes. There are many triage-based CT tools (some still under development) that detect hemorrhage and stroke (products from AIDOC, MaxQAI, VizAI), spine fractures, intra-abdominal free air and pulmonary embolism (AIDOC). To have broad relevance and adoption, offerings must expand to encompass the full spectrum of what radiologists are required to identify or exclude – all perceptible imaging findings – not to mention, assess their pertinence relative to the patient’s condition and clinical concern.

The capabilities of all-encompassing stroke-based product must include detecting large vessel occlusion, assigning ASPECTS score, identifying bleeding and mobilizing the treating
team via secure text or email. A complete product would further leverage the ability of ML to synthesize multimodal information, including advanced studies like CT angiography and perfusion imaging, with patient phenotype and suggesting the next treatment step based on predicted outcome.

Once validated, the complete offering would be essential for suspected stroke patient studies. Individual features that don’t meet the full exam requirements could still find a home consolidated in another product or perhaps on the imaging device itself.As an example, a portable X-ray unit could have the embedded capability of notifying the technologist immediately about a misplaced tube and a CT scanner could alert the technologist about urgent findings before the patient leaves the imaging suite. A product that could theoretically set aside or identify ‘normal’ on an X-ray or CT exam – the absence of actionable pathology– would have immediate and universal appeal. This could have potential impact in regions where manpower shortages prevent timely interpretation and could serve to redirect resources to reading exams with a likelihood of positive findings.  Machine learning is behind the enhancements in commercial offerings for quantitative MR and CT based neuroimaging. Initial products attacked the diagnosis of dementia (Coretechs Neuroquant) and demyelinating disease (Icometrix icoMS), an important but narrow range of diseases.

Diversification was inevitable and now both tools include the above as well as features for the assessment of epilepsy and traumatic brain injury. Ultimately, to meet the requirements of a full neuro suite and attract customers, these established companies will likely need to further expand and diversify features to encompass the broad range of neuroimaging analysis, including perfusion and perhaps fMRI via organic development or alliance with partners.

Fledgling concerns and investors in the AI space should keep these concepts in mind. Much as a feature does not make a product, a narrow application doesn’t make a company. Those who eventually provide a suite of applications that can fulfill significant, if not entire requirements of the end user – broad triage of urgent findings on a class of exams, disease based synthesis of multimodalinformation and full clinical indication based assisted diagnosis – have the best chance of thriving in this rapidly evolving field.

Lawrence N. Tanenbaum
MD FACR is VP and Chief Technology Officer and Medical Director, Eastern Region,
Radnet Inc

AIM Magazine Volume 2 Issue 3