AI radiology: market dynamics and success factors for early stage companies

Read this post to understand why we’re excited about digital health

Reach me on Twitter (Tom Moon) to discuss digital health.

Explore our State of AI 2019 Reporthighlights deck; and keynote video.

Computer vision and NLP-enabled start-ups are receiving extensive attention in radiology, given their potential to transform the speed, efficacy and scalability of diagnosis. Yet the use of machine learning (ML) is not yet common in the clinic. In order to understand the reasons behind this we’ve analysed market data and spoken with numerous vendors and healthcare payers and providers. Below we explain the drivers of adoption for hospitals, the headwinds to deployment and the success factors that start-ups need when competing with imaging incumbents, the technology giants and with each other.

Key findings…

  • ‘Disruptors’ can now unlock value at multiple points in the radiology stack including triage, diagnosis, analytics and administration.
  • Adoption is being catalysed by shortages of radiologists, national screening programmes (Europe) and pressure from payers to reduce costs (US)
  • Headwinds to adoption include the ‘CAD hangover’, high regulatory barriers, the cost of proving ROI (trials) and hospital budget dynamics
  • Successful start-ups and scale-ups must: collect diverse data sets; integrate into existing workflows; secure endorsements from key opinion leaders and ‘lighthouse’ hospitals; demonstrate compelling health economics; secure regulatory approvals; and attract significant funding
  • In a ‘red ocean’ of competition, scale-ups are attractively positioned. Select scale-ups are well-funded and could lead the field with FDA-regulated products. Incumbent equipment vendors, anticipating commoditisation, seek to transition to solution-selling. While select technology giants are increasing activity in diagnostic imaging, many focus on selling healthcare platforms and cloud infrastructure
  • Commoditisation and pricing pressure may pose challenges as deep learning proliferates. A lattice of non-technological factors can bolster defensibility
  • Key challenges for vendors to address include: fierce competition among select AI radiology scale-ups; ensuring value release within complex healthcare value chains; scaling with capital efficiency; and overcoming regulatory barriers
  • In the medium term we expect the market to rationalise to a handful of vendors focused on specific disease domains

‘Disruptors’ are unlocking value in the AI radiology imaging stack

 

Fig. 1. The AI Radiology Imaging ‘stack’

‘Impact on a clinical value chain will be a more significant value driver than position in the stack.’

Emerging technologies, such as machine learning, and workflow tools offer improved clinical outcomes and efficiency savings in the upper layers of the radiology technology ‘stack’. Above the equipment layer at the base of the stack, and the Electronic Medical Record (EMR) and Picture Archiving and Communication Systems (PACS) (Fig. 1), applications and platform are unlocking value.

Applications include AI-powered triage, decision support and diagnosis of disease sub-conditions, and natural language processing (NLP)-driven ontology modelling of radiology reports. Platforms include mobilisation and synchronisation tools for multi-disciplinary teams (MDT) and enterprise web viewers which applications can plug into.

Impact on a clinical value chain will be a more significant value driver than position in the stack. Value will accrue to vendors targeting and impacting ‘choke points’ in a clinical value chain. In the stroke pathway, time is critical; Viz.ai is scaling rapidly by delivering tools that reduce time-to-treat.

Screening programs and cost reduction will drive nascent adoption

“Nascent market adoption of machine learning-powered radiology imaging will increase.”

38% of healthcare providers intend to implement AI-enabled diagnostic interpretation within 24 months (Gartner). Nascent market adoption of machine learning-powered radiology imaging will increase due to:

  • Europe: The confluence of national screening programs (lung & breast cancer) and shortages of radiologists.
  • US: Pressure for cost savings from payers; improved outcomes and cost savings for payer-providers; and market evolution to value-based pricing.
  • Emerging markets: A desire to ‘leapfrog’ developed economies and establish screening programs and improved outcomes despite staff shortages.
  • ‘Trickle-down’ adoption: Adoption and endorsement of solutions by leading clinical institutions in respected healthcare systems, driving mainstream adoption.
  • Radiologists’ desire to better serve varying multi-disciplinary teams in their organisations.

In time, AI-powered radiology imaging solutions could become mandated for use, as were previous generations of computer-aided diagnosis despite their limited efficacy. Headwinds to AI radiology imaging include:

  • The ‘CAD hangover’: While mandated for use in US mammography, previous generations of computer-aided diagnosis increased biopsy activity without improving outcomes, causing buyer scepticism in breast imaging and beyond.
  • Proof of value: In an emerging field, evidence of efficacy is limited until validated by regulatory approvals, clinical trials and/or adoption by respected clinical institutions and key opinion leaders. For 34% of healthcare providers, trust in emerging solutions is an impediment (Gartner).
  • Budget dynamics: Allocation and alignment with changing opex, capex and reimbursement dynamics among payers

Data diversity, ‘lighthouse’ deployments, capital and regulation are key success factors

‘Effective models require training data sufficiently diversified according to geography, ethnicity, socio-economic status, morbidity and scanner manufacturer.’

Our investment framework for AI radiology highlights five vendor success factors:

1. Positioning: Leaders will offer best-in-class capability in a beachhead disease sub-condition of sufficient size, with intention to expand to other sub-conditions and imaging modalities within the disease condition

2. Data diversity: Effective models require training data sufficiently diversified according to geography, ethnicity, socio-economic status, morbidity and scanner manufacturer. Longitudinal data is additionally valuable

3. Key opinion leaders (KOLs) and ‘lighthouse’ deployments: Support from key opinion leaders, and deployments with leading institutions and programs, provide:

  • validation of product capability in a crowded field;
  • a ‘trickle-down’ effect, as additional sites adopt solutions endorsed by respected institutions;
  • ‘lock-in’, as KOLs and sites are unlikely to shift allegiance following public endorsement;
  • ‘lock-out’, as new entrants struggle to attract key stakeholders;
  • scalable revenue opportunities, given a vendor’s potential to capture a greater share of an institution’s volume, expand up the value chain (e.g. from technical recall through to diagnosis), and cross-sell solutions in other sub-conditions;
  • ‘halo’ effects, including the ability to attract additional data sets from existing partners and the opportunity to prioritise regulatory approval processes according to market priorities.

‘Leaders will offer best-in-class capability in a beachhead disease sub-condition of sufficient size’

4. Regulation: Successful vendors will achieve required CE and FDA regulations to enable market deployment, high-value use cases (diagnosis vs technical recall) and defensibility. In addition to developing products of sufficient capability, leaders will: identify appropriate use cases for regulation to avoid approvals with limited utility: prioritise the acquisition of relevant training data; recruit talent with experience in regulatory approvals to minimise timescales and costs; and attract sufficient capital to obtain multiple approvals in a minimum of time.

5. Capitalisation: AI radiology is capital intensive given the costs of: machine learning talent; data acquisition and labelling; re-training models for alternative disease sub-conditions and populations; clinical trials; CE and FDA approval processes for multiple use cases; extended pilot deployments with target clients; and 6–18 month industry sales cycles. Successful vendors will use capital as a ‘weapon’ to capture market share and defensibility through product expansion, regulatory approvals and deployments in key accounts.

In a ‘red ocean’ of competition, scale-ups are well positioned

“AI-powered radiology imaging start-ups and scale-ups are numerous. Credible vendors with appropriate success factors are fewer in number.”

Significant competitors primarily comprise equipment vendors (including GE, Philips and Siemens), technology companies (in particular, Google and Microsoft) and AI radiology start-ups and scale-ups (including Heartflow, Arterys, Viz and Zebra).

With equipment vendors attempting a transition to software/solution-selling, and global technology vendors typically focusing on large-scale infrastructure and platform deployments, AI-led radiology software scale-ups are likely to be significant vendors. Incumbent PACS vendors, such as Sectra and Carestream, are more likely to position themselves as integrators of different ML point solutions although solution vendors will try to resist this.

Equipment vendors are transitioning to solution-selling

Equipment manufacturers anticipate the commoditisation of equipment and seek to transition to solution-selling incorporating software components. Analogously to their strategy in the manufacturing sector, for predictive maintenance, equipment vendors will access radiology software solutions through development, go-to-market partnerships and acquisition.

Equipment vendors can leverage their hardware assets, relationships, distribution and resources.

But they must navigate significant challenges. These include: navigating a business model transition from hardware sales to solution-selling; translating equipment advantage to data acquisition; longer cycles of innovation; and a desire for vendor lock-in that limits suitability in heterogenous environments.

Among global technology companies, Google’s capability is growing

Global technology companies, including Microsoft, offer software infrastructure and platforms for healthcare institutions, while Google is developing deep-learning based diagnostic solutions for imaging.

Technology companies are taking advantage of their strengths. Microsoft Azure is the only NHS-approved cloud environment, offering significant market penetration. Google DeepMind has world-class deep learning expertise, initial partnerships (including the Royal Free for acute kidney injury, and Moorfields for diabetic retinopathy screening), extensive resources and sees a range of opportunities in the healthcare sector.

However, to the benefit of radiology start-ups and scale-ups, Microsoft is focused on platform opportunities (large-scale software infrastructure engagements and Azure deployment) to a greater extent than application development. And, via its Streams solution, Google DeepMind is prioritising the productisation and commercialisation of data amalgamation and clinician synchronisation tools.

However, DeepMind is evolving its capability and activity in image recognition, with recent progress in lung CT diagnosis in particular, and remains a significant longer-term threat to earlier stage companies.

Competition is intense among imaging start-ups and scale-ups

AI-powered radiology imaging start-ups and scale-ups are numerous. Credible vendors with appropriate success factors are fewer in number. Several focus on adjacent areas or have a different focus to expansion within a disease condition. Select vendors have the potential for significant presence given material funding, multiple products, regulatory approvals and a growing set of deployment relationships.

There are several exciting, early stage start-ups focused on various aspects of ML-enabled radiology. Examples of select, leading scale ups with FDA approved, commercial products include:

  • Heartflow benefits from: a strong focus on a single disease condition (coronary heart disease); a leading, FDA-approved product, reimbursement from Medicare, Medicaid and multiple US payers; and significant scale in personnel and funding.
  • Arterys is expanding from its initial focus on cardiac MRI to offer a broader platform into which third party algorithms can integrate. Arterys benefits from: multiple point solutions; a ‘horizontal’ enterprise solution (web viewer) open to third parties; and significant scale, funding and international presence.
  • Viz.AI is focused on care synchronisation to address stroke, rather than ‘AI radiology’. Its strengths include: a suite of tools (including a workflow solution, mobile PACS viewer, and HIPPA-compliant-messaging tool) that reduce the time to thrombolysis; strong ROI for multiple stakeholder groups; de novo FDA approval for CT head stroke triage; and significant funding from respected investors.
  • Zebra offers multiple point solutions for disease specialities, with varying value propositions (mammography screening, compression fractures, acute intracranial haemorrhage triage). Zebra benefits from: a strategic investor relationship with InterMountain Healthcare; significant scale and substantial funding.

Market dynamics will bolster defensibility

“A lattice of non-technological factors can bolster defensibility.”

The proliferation of deep learning may pose challenges of commoditisation and pricing pressure.

In practice, a lattice of non-technological factors can bolster defensibility, which will come by degree. Requirements for domain expertise, diverse data, regulatory approvals, and endorsements from respected institutions and healthcare systems will limit commoditisation.

Further, market-leading technology will enable marginal gains in capability that unlock additional, defensible areas higher in the value chain — such as diagnosis.

Competition and capital efficiency are key risks for start-ups to mitigate

The AI radiology market presents an attractive opportunity for value creation given:

  • the potential for AI-powered radiology solutions to unlock improved clinical outcomes and cost savings;
  • a US payer environment moving to value-based outcomes over the medium– to long-term;
  • the opportunity to establish leadership in a beachhead, expand within a disease condition and multiple imaging modalities, and explore platform-based opportunities within the disease condition pathway;

Key challenges to mitigate are:

  • fierce competition between select AI radiology scale-ups, and from vendors in adjacent areas of the radiology technology stack;
  • ensuring utility unlocks value within complex healthcare value chains;
  • scaling with capital efficiency across disease sub-conditions and imaging modalities
  • overcoming regulatory and pragmatic barriers to deployment

We expect market rationalisation

In the medium term we expect the market to rationalise to a select number of vendors focused on specific disease domains, given that:

  • healthcare providers do not have the capacity or inclination to engage with multiple point solution vendors
  • the difficulty of accessing training data, integrating with workflows, and building trust with end users favours specialising around disease domains such as cardiovascular, oncology and neurology.
  • only a select number of start-ups will achieve scale, given the multiple, challenging success factors required to bring a product to market at scale.

 

Want to discuss this article? Reach me on Twitter Tom Moon

Want to know more about MMC’s healthcare focus? Read Tom’s post on investing in the future of healthcare

Many thanks to David Kelnar. This post originally appeared on MMC Writes

Back to Research

Portfolio

Read more...

You Might Be Interested In
  

Team

Read more...

You Might Be Interested In
  

Research

Read more...

You Might Be Interested In
  

News

Read more...

You Might Be Interested In
  
Request For Information
Important notice