From X-Rays to AI: Navigating US Regulations in Radiological Health
115
as a tool in the clinical environment, widespread
adoption in routine clinical practice faces some
obstacles including trustworthiness, reliability,
transparency, explainability, responsibility, privacy
and security.8
This chapter provides an overview of AI/ML
functionalities used for imaging and describes
the evolution of FDA regulations related to AI/
ML software in the radiological health space.
Throughout this chapter the terms AI/ML soft-
ware or technologies or systems or applications
are used interchangeably. This chapter does not
cover the standalone software used for auto-con-
touring in nuclear medicine and radiation
therapy (NMRT) applications. For more infor-
mation on NMRT software devices please refer
to Chapter 6. This chapter also offers some key
considerations for SaMD developers to ensure
that their technology is safe and effective when
deployed for clinical use. Finally, it provides an
outlook on where the field is headed through the
lens of the authors. A summary of the key terms,
abbreviations and definitions used throughout
this chapter is provided in Table 8-1.9-22
Historic Perspectives
The clinical use of software tools incorporating
AI/ML is becoming more widespread within
the radiological health space. The software tools
have a range of applications, including assisting
healthcare providers in:
• Image acquisition
• Image reconstruction
• Image processing (e.g., from computed
tomography [CT], magnetic resonance
[MR], ultrasound, positron emission tomog-
raphy [PET], and single-photon emission
computerized tomography scans)
• Image management and storage
• Automation of radiology workflows via a
radiology information system (RIS), and
• Detection, segmentation, classification, and
quantification of pathological findings.
Within radiology, computer-assisted devices
(CAD) are becoming widely available in clinics
to improve radiology workflow, with particular
applications to the triage (CADt), detection
(CADe) and diagnosis (CADx), or both (CADe/x)
of different abnormalities in medical images.
Early research and development efforts to
analyze medical images with CAD started in the
1960s.23 More intense research investigation of
CAD began in 1980s with the use of improved
automation tools for conducting computer analy-
sis on lesions seen on medical images.24 The FDA
approved the first CAD software in 1998 via the
premarket approval (PMA) process to help radiolo-
gists identify and mark regions of interest (ROIs) on
routine screening mammograms. Since then, numer-
ous CAD devices have reached the US market,
with the rate of marketing authorizations having
increased substantially since 2016 (Figure 8-2).25-33
FDA-cleared or Approved AI
Applications in Radiological Health
The FDA regulates software devices based on
their risk to the patient (as previously described
in Chapter 2, which provided an overview of reg-
ulatory pathways relevant to medical devices, and
also in Chapter 7, which provided an overview of
software regulation).
Available regulatory pathways for AI/ML
enabled medical devices include:
1. The PMA pathway – the most stringent
review for high-risk devices)
2. The 510(k) premarket notification pathway-
for moderate-risk devices and
3. The De Novo premarket review (for novel
low- and moderate-risk devices).
Based on publicly available data on FDA
marketing authorizations, the FDA has cleared
96.4% of AI/ ML medical devices via the 510(k)
pathway, granted Class II marketing authori-
zation to 3.5% via the De Novo pathway, and
approved the rest via the PMA process.6
While AI/ML algorithms play a role in the
operation of many radiological health devices,
the rest of this chapter will focus on AI/ ML
image processing devices as well as CADe,
CADx and CADt devices. Figure 8-3 provides
examples of CAD applications cleared in recent
years within the radiology field.
115
as a tool in the clinical environment, widespread
adoption in routine clinical practice faces some
obstacles including trustworthiness, reliability,
transparency, explainability, responsibility, privacy
and security.8
This chapter provides an overview of AI/ML
functionalities used for imaging and describes
the evolution of FDA regulations related to AI/
ML software in the radiological health space.
Throughout this chapter the terms AI/ML soft-
ware or technologies or systems or applications
are used interchangeably. This chapter does not
cover the standalone software used for auto-con-
touring in nuclear medicine and radiation
therapy (NMRT) applications. For more infor-
mation on NMRT software devices please refer
to Chapter 6. This chapter also offers some key
considerations for SaMD developers to ensure
that their technology is safe and effective when
deployed for clinical use. Finally, it provides an
outlook on where the field is headed through the
lens of the authors. A summary of the key terms,
abbreviations and definitions used throughout
this chapter is provided in Table 8-1.9-22
Historic Perspectives
The clinical use of software tools incorporating
AI/ML is becoming more widespread within
the radiological health space. The software tools
have a range of applications, including assisting
healthcare providers in:
• Image acquisition
• Image reconstruction
• Image processing (e.g., from computed
tomography [CT], magnetic resonance
[MR], ultrasound, positron emission tomog-
raphy [PET], and single-photon emission
computerized tomography scans)
• Image management and storage
• Automation of radiology workflows via a
radiology information system (RIS), and
• Detection, segmentation, classification, and
quantification of pathological findings.
Within radiology, computer-assisted devices
(CAD) are becoming widely available in clinics
to improve radiology workflow, with particular
applications to the triage (CADt), detection
(CADe) and diagnosis (CADx), or both (CADe/x)
of different abnormalities in medical images.
Early research and development efforts to
analyze medical images with CAD started in the
1960s.23 More intense research investigation of
CAD began in 1980s with the use of improved
automation tools for conducting computer analy-
sis on lesions seen on medical images.24 The FDA
approved the first CAD software in 1998 via the
premarket approval (PMA) process to help radiolo-
gists identify and mark regions of interest (ROIs) on
routine screening mammograms. Since then, numer-
ous CAD devices have reached the US market,
with the rate of marketing authorizations having
increased substantially since 2016 (Figure 8-2).25-33
FDA-cleared or Approved AI
Applications in Radiological Health
The FDA regulates software devices based on
their risk to the patient (as previously described
in Chapter 2, which provided an overview of reg-
ulatory pathways relevant to medical devices, and
also in Chapter 7, which provided an overview of
software regulation).
Available regulatory pathways for AI/ML
enabled medical devices include:
1. The PMA pathway – the most stringent
review for high-risk devices)
2. The 510(k) premarket notification pathway-
for moderate-risk devices and
3. The De Novo premarket review (for novel
low- and moderate-risk devices).
Based on publicly available data on FDA
marketing authorizations, the FDA has cleared
96.4% of AI/ ML medical devices via the 510(k)
pathway, granted Class II marketing authori-
zation to 3.5% via the De Novo pathway, and
approved the rest via the PMA process.6
While AI/ML algorithms play a role in the
operation of many radiological health devices,
the rest of this chapter will focus on AI/ ML
image processing devices as well as CADe,
CADx and CADt devices. Figure 8-3 provides
examples of CAD applications cleared in recent
years within the radiology field.