DefenceDX AI is an innovative machine learning (ML)-based prototype (proof of concept) for the classification of medical image data. With its evaluations, it offers a quick decision aid to the treating doctors. Its first area of application is the detection of pneumonia on X-ray images.
How it works
Using DefenceDX AI is as simple as can be. After uploading a lung X-ray in JPG or PNG format, it will be automatically analysed by the system. The artificial intelligence (AI) is trained to recognise typical inflammation patterns and delivers a twofold result in the blink of an eye:
- Presence of pneumonia: yes / no
- If yes: bacterial or viral infection
Both results are displayed with a relative probability. They make it easier for the doctor to make quick and informed decisions about further treatment.
In a later stage of development, the system will assess the presence of various lung infectious diseases, e.g. SARS-CoV-2, tuberculosis or inflammatory lung diseases.
Our DefenceDX AI is a web-based system that can be operated in public clouds (web servers) or in closed intranet environments. Both PCs and mobile units can serve as end devices. Due to the low technical requirements, DefenceDX AI is ideally suited for use in medical practices and hospitals that do not have access to an expensive computer tomograph (e.g. in developing countries).
The machine learning algorithm on which the system is based can also be used for other X-ray-enabled indications, such as malignant neoplasms or metastases in the lungs, pulmonary congestion, pulmonary effusion formation, cardiac insufficiency or pneumothorax. Furthermore, an implementation of additional classification tasks is possible, as well as the integration of further medical parameters (e.g. laboratory values or other medical course parameters).
We would be happy to develop solutions for your special application needs. Please do not hesitate to contact us.
If you are interested in DefenceDX AI, please contact:
Alpspitz Bioscience GmbH
Phone: +49 (0) 9281-84016-260
A deep learning based assistance system for the classification of pneumonia using X-ray images
Master thesis by Jan Raber, 2021
In view of the given dataset, the thesis by Jan Raber investigates the current limitations of the prototype, which classifies CXR images according to pneumonia yes/no (stage 1), bacterial/viral (stage 2). Moreover a visualization component for this prototype has been implemented.
Evaluation of ensemble learning for disease classification on chest radiographs
Master thesis by Sebastian Steindl, 2021
With respect to the Machine Learning approach used, the thesis by Sebastian Steindl examines the limitations of this prototype in the algorithmic classification itself. Here, the focus is not on the data set itself, but on the steps (which extend to a re-implementation of the model) that are required to develop the prototype towards a pilot that allows a more reliable classification. To this end, methods from recent publications were analysed, implemented and combined (using ensemble learning, among others) to perform a multi-label classification on the CheXpert dataset that also takes into account uncertainties from radiology reports. The final ensemble achieved an average ROC-AUC of 0.917 for the five relevant labels of the hidden test set provided by the CheXpert competition.