Chapter 8. Artificial Intelligence-based Software
120 Regulatory Affairs Professionals Society (RAPS)
the necessary tests and performance criteria for
display devices used in medical imaging to ensure
consistent presentation of medical images.37
Common performance test metrics MIMPS
include spatial resolution, measurement accuracy,
segmentation accuracy, reproducibility, signal
to noise ratio (SNR), root mean square error
(RMSE), Dice Similarity Coefficient (DSC),
the Hausdorff distance, among others. It is the
manufacturers responsibility to demonstrate the
standalone performance tests meets or exceeds
the preset acceptance criteria defined in the stan-
dard and task group report to ensure the safety
and effectiveness of these devices.
One example of an image processing device
includes AI-Rad Companion (Pulmonary)
(K213713),38 an image-processing software that
provides quantitative and qualitative analysis
from previously acquired CT DICOM images to
support radiologists and physicians in the evalua-
tion and assessment of disease of the lungs.
Medical Image Communication Systems,
Medical Device Data System, and
Medical Image Storage Device – 21 CFR
§892.2020, §880.6310, and §892.2010
Medical imaging communication, data and
image storage regulation is covered by 21 CFR
§892.2020, §880.630, and §892.2010.
A sub-set of AI tools is intended to augment
workflow solutions by reducing burden of basic
repetitive tasks and increasing diagnostic pre-
cision when interpreting medical images. These
are clinical decision software devices that may be
considered Class I medical devices if the software
functions are not performing advanced image
processing functions.
Class I medical devices are products deemed
to be low-risk, and as such are subject to the least
amount of regulatory control including exemp-
tion from pre-market notification.
There are several AI/ML software solutions
in this category that can be fully integrated into
the image interpretation workflow, and which are
intended to help clinical users handle their daily
workload. Two examples include the multi-mo-
dality viewer (K182230)39 and the Hologic
SecureView DX-RT (K041555).40
CADe, CADx, and CADt Devices
Software incorporating AI/ML has applications
in various types of clinical decision support. Here,
the discussion is limited to CADe, CADx, and
CADt applications, as overviewed in Figure 8-4.41
CADe (21 CFR §892.2070): Medical Image
Analyzer
CADe systems use algorithms to recognize
patterns on radiological images to detect
abnormalities, such as a tumor or other lesion
indicative of a disease condition. The majority of
CAD devices introduced on to the US market
between 1998 and 2016 were CADe systems.23
These devices cover a wide range of clinical
applications from highlighting suspicious areas
of tissue on mammography or chest x-rays
to improving the detection of lung nodules.
Examples of CADe devices include Chest-CAD
(K210666)42 that analyzes chest radiographs
using machine learning techniques to identify,
categorize, and highlight suspicious ROIs and
auto lung nodule detection (K201560)43 which is
used to identify and mark regions of suspected
pulmonary nodules.
Several research studies suggest that CADe
can improve a radiologist’s ability to detect breast
abnormalities on mammograms with increased
confidence.44 After the first CADe device became
available on the US market in 1998, CADe adop-
tion was initially slow but then began to increase
dramatically in digital mammography, jumping
from 5% in 2003 to 83% in 2012.44
When collating data to support FDA
submissions of new CADe and CADx devices,
clinical study design and the statistical methods
used are important to consider. A successful
submission should include both standalone
performance testing ,which is testing that
demonstrates the device improves performance
in the intended use population when used in
accordance with the instructions for use (IFU)
labeling, and a reader study. Manufacturers
typically report details of the device’s sensitivity
and specificity, receiver operating characteristic
(ROC) curves, and areas under the ROC curves
(AUCs).45 Setting an appropriate objective
reference standard (ground truthing process) and
120 Regulatory Affairs Professionals Society (RAPS)
the necessary tests and performance criteria for
display devices used in medical imaging to ensure
consistent presentation of medical images.37
Common performance test metrics MIMPS
include spatial resolution, measurement accuracy,
segmentation accuracy, reproducibility, signal
to noise ratio (SNR), root mean square error
(RMSE), Dice Similarity Coefficient (DSC),
the Hausdorff distance, among others. It is the
manufacturers responsibility to demonstrate the
standalone performance tests meets or exceeds
the preset acceptance criteria defined in the stan-
dard and task group report to ensure the safety
and effectiveness of these devices.
One example of an image processing device
includes AI-Rad Companion (Pulmonary)
(K213713),38 an image-processing software that
provides quantitative and qualitative analysis
from previously acquired CT DICOM images to
support radiologists and physicians in the evalua-
tion and assessment of disease of the lungs.
Medical Image Communication Systems,
Medical Device Data System, and
Medical Image Storage Device – 21 CFR
§892.2020, §880.6310, and §892.2010
Medical imaging communication, data and
image storage regulation is covered by 21 CFR
§892.2020, §880.630, and §892.2010.
A sub-set of AI tools is intended to augment
workflow solutions by reducing burden of basic
repetitive tasks and increasing diagnostic pre-
cision when interpreting medical images. These
are clinical decision software devices that may be
considered Class I medical devices if the software
functions are not performing advanced image
processing functions.
Class I medical devices are products deemed
to be low-risk, and as such are subject to the least
amount of regulatory control including exemp-
tion from pre-market notification.
There are several AI/ML software solutions
in this category that can be fully integrated into
the image interpretation workflow, and which are
intended to help clinical users handle their daily
workload. Two examples include the multi-mo-
dality viewer (K182230)39 and the Hologic
SecureView DX-RT (K041555).40
CADe, CADx, and CADt Devices
Software incorporating AI/ML has applications
in various types of clinical decision support. Here,
the discussion is limited to CADe, CADx, and
CADt applications, as overviewed in Figure 8-4.41
CADe (21 CFR §892.2070): Medical Image
Analyzer
CADe systems use algorithms to recognize
patterns on radiological images to detect
abnormalities, such as a tumor or other lesion
indicative of a disease condition. The majority of
CAD devices introduced on to the US market
between 1998 and 2016 were CADe systems.23
These devices cover a wide range of clinical
applications from highlighting suspicious areas
of tissue on mammography or chest x-rays
to improving the detection of lung nodules.
Examples of CADe devices include Chest-CAD
(K210666)42 that analyzes chest radiographs
using machine learning techniques to identify,
categorize, and highlight suspicious ROIs and
auto lung nodule detection (K201560)43 which is
used to identify and mark regions of suspected
pulmonary nodules.
Several research studies suggest that CADe
can improve a radiologist’s ability to detect breast
abnormalities on mammograms with increased
confidence.44 After the first CADe device became
available on the US market in 1998, CADe adop-
tion was initially slow but then began to increase
dramatically in digital mammography, jumping
from 5% in 2003 to 83% in 2012.44
When collating data to support FDA
submissions of new CADe and CADx devices,
clinical study design and the statistical methods
used are important to consider. A successful
submission should include both standalone
performance testing ,which is testing that
demonstrates the device improves performance
in the intended use population when used in
accordance with the instructions for use (IFU)
labeling, and a reader study. Manufacturers
typically report details of the device’s sensitivity
and specificity, receiver operating characteristic
(ROC) curves, and areas under the ROC curves
(AUCs).45 Setting an appropriate objective
reference standard (ground truthing process) and