49th Union World Conference on Lung Health , 24 - 27 October, 2018, The Hague
At the 2018 UNION Conference in The Hague, two Symposia on digital radiology and artificial intelligence such as Computer Aided Detection for Tuberculosis are organised. Evidence on high e.g. CAD4TB sensitivity and specificity scores on data sets, from e.g. Pakistan and South Africa, are highlighted at the Conference. Also data on the use of CAD4TB for detecting TB in children will be presented.
Session Type: Satellite session UNION2018
Session Track: A - Advances in Quality of TB Care and Services
Role of artificial intelligence and chest radiography for the detection of tuberculosis and other lung diseases
Digital chest radiography is increasingly used to screen for tuberculosis. Used as a triage test to select individuals for bacteriological testing, it allows for screening large groups of TB suspects at low costs. Chest radiography is also essential for the diagnosis and monitoring of many other lung diseases. Artificial intelligence is a rapidly developing field and increasingly applied to medical image analysis. The first products for chest radiography analysis have appeared on the market. In this symposium we show and discuss progress and application of chest radiography and artificial intelligence to the detection of TB and other lung diseases.
Target Audience 1: TB physicians
Target Audience 2: Pediatric physiciansb
Target Audience 3: Pulmonologistsb
Target Audience 4: Radiologistsb
Target Audience 5: Researchers
Session sponsor: Delft Imaging Systems, firstname.lastname@example.org, +31 639 672 569
Chair 1: Bram van Ginneken, Radboudumc, Nijmegen, Netherlands
Chair 2: Klaus Reither, Swiss Tropical and Public Health Institute, Basel, Switzerland
Shifa S. Habib (Karachi, Pakistan)
Scale-up of computer-aided detection for TB in Pakistan: Implementation experience and development of individual risk scoring algorithms.
In Pakistan, the advent of CAD4TB has greatly facilitated mass-screening through automated interpretation of chest X-rays without the need of onsite trained radiologists or clinicians. Currently, in Pakistan over 50 mobile CAD4TB supported mobile X-ray units are being rolled out. This presentation discusses the operational considerations in CAD4TB scale-up that may limit the effectiveness of case finding programs, such as selecting optimum threshold scores or obtaining viable sputum samples from high-score patients for testing. Early-stage research from our group has shown that risk of TB increases significantly for CAD4TB scores >80. In addition, variations in demographics and clinical history can generate different individual risk probabilities at the same score. In this presentation we discuss how these probabilities may be utilized to develop an algorithm to determine personalized individual risk scores for every visitor at active case finding programs.
Bram van Ginneken (Nijmegen, Netherlands)
Artificial Intelligence and Deep Learning in Chest Radiography: Automating the Reading Process.
Artificial intelligence, machine learning, and deep learning are rapidly developing technologies for medical image analysis. They have the potential to improve access and reduce costs for chest radiography, particularly in countries where there is a shortage of trained human readers. In this talk, I will provide an introduction and an overview of these techniques with a focus on chest radiography. I will show results on automated processing of chest x-rays for early detection, monitoring and quantification of diseases such as tuberculosis, but also silicosis, pneumonia, emphysema and lung cancer. Recent results will be presented on the sensitivity and specificity of automated TB detection compared to human expert reading and bacteriological testing, with and without deep learning, on large data sets from several countries in Africa and Asia.
Nasreen Mahomed (Johannesburg, South Africa)
Computer Aided Diagnosis for WHO standardized Chest X-ray interpretation in Children.
There are a limited number of paediatric radiologists in low income countries, including those with experience and certification in WHO standardized chest X-ray interpretation in children. The application of computer aided diagnosis (CAD) in chest X-rays of children may be an interesting alternative and supplement to human reading. Pneumonia is among the leading infectious causes of morbidity and mortality in children under five years old globally. We evaluated several convolutional neural network architectures for the automated interpretation of chest X-rays for the presence of primary endpoint pneumonia. We used data from hospitalized South African children with pneumonia participating in the PERCH and PCV-13 studies. On an independent test set of 600 cases, an independent human observer achieved an AUC of 0.833. The best deep learning network ensemble achieved a significantly better performance of 0.876. The results show there is great potential for computer-assisted paediatric CXR reading using artificial intelligence.
Anja van't Hoog (Amsterdam, Netherlands)
Choosing algorithms for TB screening: triage tests with chest radiography, symptoms, and the Xpert® MTB/RIF assay.
The choice of an appropriate screening and diagnostic algorithm for tuberculosis (TB) screening depends on the setting. In this talk I will discuss various algorithms composed of currently available methods, including symptom screening, chest radiography, sputum-smear microscopy, and Xpert® MTB/RIF as confirmatory test. The ideal algorithm does not exist. The choice will be setting specific, for which guidance can be provided.
SELECTED SESSION TYPE: Satellite session.
Session sponsor: Qure.ai
Technological advances in digital x-ray systems along with the emergence of high performance computer aided diagnosis (CAD) software platforms offering real time point-of-care results have led to renewed interest in the possibility of deploying CXR, even in resource scarce settings, as the primary screening/triage test for PTB and indeed the spectrum of lung disease - chronic & infectious.
Newer affordable portable digital x-ray systems designed for field use in low resource settings have resulted in an increased number of active case finding programs and prevalence surveys. Results indicate that when interpreted consistently, CXR is the most sensitive screening tool for pulmonary TB and that a significant proportion of people with TB are asymptomatic, at least early in the course of the disease. CAD based CXR readings can be expected to minimize intra-/inter- reader variability, allowing consistent rule-based interpretation.
The effective use of CXR as the primary screening test (over microscopy) for suspect candidates recommended by the primary caregiver, with candidates with abnormalities referred for Xpert® for confirmatory/susceptibility testing has significantly contributed to the landmark success of PATH’s Mumbai PPIA & IRD’s Karachi screening and active case finding programs.
From an integrated health systems and rights-based universal access to high quality diagnostics perspectives – the above diagnostic algorithm is easily extended to include confirmatory testing for pneumonia and other pathologies as well as for chronic lung diseases.
Furthering the agenda of previous CXR symposiums in Cape Town (2015) & Liverpool (2016) – the symposia will focus on highlighting the present and future of CXR radiology. A panel of expert stakeholders will present their ongoing experience in deploying X-ray hardware and CAD software and discuss the mandate for the future of CXR as a primary screening test for lung symptomatics in the context of UHC and TB 2025 agenda.
18:00 – 18:10: Advances in digital radiology hardware for resource limited settings Kunal Bose, India
18:10 – 18:20: Automated Chest X-ray reads with deep learning Prashant Warier, India
18:20 – 18:35: Large scale active case finding using automated CXR: The IRD experienceAamir Khan, Pakistan
18:35 – 18:50: CXRs as primary triage: The Mumbai PPIA model Shibu Vijayan, India
18:50 – 19:05: Early adopter experiences in deploying automated CXR reading in TB centers accross DhakaSayera Banu, Bangladesh
19:05 – 19:20: The road ahead for automated CXR Jacob Creswell, Switzerland
19:20 – 19:30: Discussion