The CAD4TB AI algorithm allows TB programs to use the software offline on a laptop computer and/or in the cloud. In resource constrained areas where internet connectivity is often still unstable or expensive per MB, the offline version of CAD4TB can be the only workable option.

For those settings whereby internet connectivity is stable with an adequate bandwidth and low data bundle cost the Cloud solution can be an alternative for offline use of CAD. The uploading of the automatically compacted dCXR image to the CAD4TB AI algorithm in the cloud normally takes 30-40 seconds and the CAD outputs will be available on the web to the user on site and/or another location anywhere in the World in about 10-15 seconds.

Offline CAD

The algorithm outputs are the heat map and CAD abnormality score within about 30 seconds on the laptop that is connected to the digital X-ray system on site. This is also of special operational value in case of outreach screening programs to remote places with poor or no internet connectivity.

• Offline solution, no need for internet connectivity
• Images will be stored locally and may require offline image back-up provisions
• Images can be synchronized with CAD4TB cloud when internet is available
• No internet connectivity or hosting cost

Cloud & local Data Server CAD

For those settings whereby internet connectivity is stable with an adequate bandwidth and low data bundle cost the Cloud solution can be an alternative for offline use of CAD. So, uploading the automatically compacted DICOM image to the CAD AI algorithm in the cloud takes about 30-40 seconds, while the CAD outputs will be available on the web to the authorised user on site and/or another location anywhere in the World in about 10-15 seconds.

For the Cloud CAD set-up images and data will be stored online; data back-up can be automated.
However, some countries prohibit cloud storage of personal medical data outside their country for such settings another solution may be considered:

• Using a CAD4TBServer set-up in-country can be a way to store medical data in compliance with local data protection legislation with an in-country data server;
• CAD4TBServer can serve as the local central storage hub for the various in-country users;
• Safe VPN connects all CAD4TBBox user sites with the in-country CAD4TBServer
• All TB data like, DICOM images, patient information, CAD4TB output and lab results from multiple sites can this way be stored and accessed centrally in-country

CAD Input-Output

Input

• Posterior Anterior Chest X-ray (PA CXR)
• Subject’s age: 4 years or above

Output

• Quality assessment of the digital CXR
• Abnormality score: 0 -100 + heatmap indicating areas with abnormalities
• Higher CAD score higher % positive Xpert® tests

Diagnostic use of CAD

CAD4TB used in automated pre-screening significantly reduces cost and time to diagnose TB for patient and provider [1]. Users can take the above outputs into account in their clinical work: they can decide that the subject should undergo further testing for the presence of TB or other lung diseases in case the heat map displays suspicious regions, verified by a human operator as suspicious, or when the CAD score is above a certain threshold. The user’s choice for this CAD abnormality threshold will depend on the settings in which the software is used. There is a trade-off between sensitivity and specificity (higher sensitivity often means lower specificity) which can also depend on the available TB program budget. So if the budget is limited and the risk group large then selecting a higher CAD score threshold, resulting in less false positives, can allow the program to detect more TB cases, while accepting that some cases will be missed. CAD in addition may alleviate other issues related to CXR interpretation, such as high intra- and inter-observer variability [2].

CAD as TB risk “traffic light”

So in practice when used as a triage before Xpert®: a presumptive TB patient could for instance be defined as anyone who has a CAD abnormality score of >65 when screened using dCXR. Using CAD in combination with age and gender data enhanced the performance of the software; variations in demographic information can generate different individual risk probabilities for the same CAD4TB scores [3]. The combination of CAD and clinical information can offer improved accuracy and increased specificity compared with using either type of information on its own [4]. In various studies the aptitude of dCXR to detect pre-clinical TB was confirmed. This could also allow TB programs to use CAD as a kind of TB risk traffic light with the following sample thresholds for abnormal, subnormal and normal:

---