Radboud University of Nijmegen developes the Computer Aided Detection for TB diagnostic software in close cooperation with Rogan, Zambart and the Lung Institute of Cape Town. CAD is already as accurate as trained human readers abnormalities consistent with TB. Partial financing for this scientific research was obtained from the Dutch Ministry of Economic Affairs with support from CheckTB!.
The CAD system follows the general machine learning approach of supervised learning. In supervised learning the system learns from a set of examples, called the training data. The training data consists of a number of labelled images of which the TB status is known. For the CAD4TB system there are two classes and the labels are normal and suspect for TB. A number of features is extracted from these images and combined to form a feature vector for each image. These feature vectors are used to train a TB classifier. This TB classifier incorporating state-of-the-art pattern recognition techniques can then be used to classify images of which the label is not known.
Computer Aided Diagnosis will enable automated accurate pre-screening on site by the digital X-ray system allowing radiologists or pulmonologists to focus more on analysing the TB suspect images. In addition CAD4TB will contribute to fewer "over-reading" by less experienced radiographers or clinical officers. CAD in combination with molecular tests such as Xpert innovates screening for early case detection. This CAD-Xpert combination enables rapid and more cost effective active case finding in high risk groups.
Possibilities for Future Use CAD4TB
To increase prevalence survey accuracy and to reduce cost of risk group screening programs CAD offers the following diagnostic functionality:
CAD4TB also allows for pre-selecting its sensitivity, which is of special value for prevalence surveys. Once the CAD4TB protocol is optimised it is expected the number of CXR requiring human reading can be reduced to less than 20%. For the moment in upcoming prevalence surveys where CAD4TB will be used dual reading (CAD and human) will done.
Basis for the revolution in diagnostics
In 2004, the Image Science Institute of the Medical Centre of Utrecht University and Rogan (both of the Netherlands) prepared a prototype software protocol able to recognise TB suspects in a database of 500 images with a sensitivity of 85% and a specificity of 50%. Valuable experience gained over the years from the development of accurate Computer Aided Detection of breast cancer, supporting mammography examinations, was used from the start.
The digital image can soon (last quarter of 2010) be interpreted on the spot within seconds by the CAD4TB protocol programmed to recognise image patterns consistent with TB. So digital Chest images can be read on the spot in real time by the CAD protocol, by the radiographer on a monitor or sent over any mobile phone or internet to any venue for second reading by a distant radiologist.
These are breakthrough new developments in boosting TB case detection using CXR and are welcomed by many institutions in Africa and leading TB organizations in the World like the UNION, WHO and the Dutch KNCV TB Foundation.
The “Computer Aided Detection for TB (CAD4TB)” project is in full development. The University of Nijmegen is now involved on the Dutch side together with Rogan they collaborate with the University of Cape Town and the Desmond Tutu Lung Institute of the University of Stellenbosch (both of South Africa) and Zambart (Zambia) to further develop and scientifically proof this breakthrough diagnostic capability.
The objective of the joint research is to achieve an electronic TB screening capability with a sensitivity of 90% and a specificity of 80%. The specificity target will be the biggest challenge for the scientists. It is expected that on the basis of reading thousands of images (including access to patient information like HIV status), that patterns consistent with TB typical for co-infected patients will be found. This will further enhance the role of X-ray to effectively detect TB suspects in people living with HIV.
CAD improves TB screening and supports active case finding by:
A high percentage of TB suspect cases is identified easily wherever the digital camera is used.
- Reducing number of human image interpretations or when used as second opinion reducing under and over reading;
- Providing a powerful tool in prevalence studies;
- Offering an effective quality assessment tool for health workers and NTP;
- Cutting screening and diagnostics costs.
The reference images (including data on smear microscopy and culture) available in the database will support radiographers to interpret Chest images on the spot with higher accuracy.
Only a limited number of complex Chest images will require external radiological expertise.
While developed for TB, this system can be enhanced in the future to spot other lung diseases as well. It is thus expected to contribute to a broad Healthcare strengthening impact in an area where this is most needed.
A small scale study conducted in Zambia (2009) clearly shows the potential such a CAD system has. This preliminary study compares the performance of a prototype CAD system with human readers in classifying radiographs. Six clinical officers were given a test set of 157 images which they labelled normal or suspected to TB. Gold standard of these 157 images was the expert diagnosis of two experienced Dutch radiologists who labelled the images normal and suspected to TB. With more images collected every day including patient information on HIV status, sputum microscopy and culture results, the performance of the system will improve considerably over time.
A number of 60 disease-free and 97 images suspect for TB were presented blinded to 6 human readers and to the CAD system. The readers were clinical officers. The readers indicated for each image whether it was normal or suspect for TB. The CAD system was initially trained with another set of images and assigned a probability of being suspect for TB based on the textural appearance of the images. The sensitivity, specificity and accuracy of the readers and of the CAD system were subsequently determined.
The table shows the results for the readers and the CAD system. The average sensitivity and specificity for the readers were respectively 0.68, 0.49 and 0.60, but the variability in reader performance was high. Reader 4 performed similar to the CAD system, but the other readers performed less well. At the same specificity as the average of the readers the CAD system would obtain an sensitivity of 87%.
Digital chest images are forwarded daily to the scientific database for ongoing CAD4TB development from UTH, Kanyama Clinic and Lung Institute of Cape Town to further increase sensitivity and specificity levels. From the small scale Zambia study it can be concluded, that the CAD system performed similar or better than the human readers in classifying radiographs as being normal or suspect for TB. However, false detection rates at a reasonable sensitivity are high for both. The performance of both human readers and the CAD system is expected to improve significantly with more training.
Can your organization participate in the study?