AI-based solution for Covid-19 pneumonia

Our Mission

FastRAI’s goal is to develop an AI-based solution to retrieve more information from plain thoracic X-rays regarding Covid-19 pneumonia and progression of the disease.


Immediate Goal


We want to support clinicians in hospitals overburdened by Covid-19 cases. Therefore, we offer an opportunity to upload chest CT data and get a “second reading” in return which has been done by a board certified radiologist.

Online Portal

deepc and M3i take care of this first step by creating an online portal for uploading and managing studies, by providing a safe and usable solution for proper data anonymization and by routing the datasets to our radiologists who will then create a structured report. The clinicians will then receive the results on their online worklist.

AI Models

Moreover, with time we will deploy AI models as well so we can provide not only an expert report on the case but segmentation, volumetrics, classifications of disease and more.


Intermediate Goal (TUM)

FastRAi aims to develop neural network based applications for X-ray based CoVid diagnosis. The applications are developed using labelled CT data. Neural networks are powerful artificial intelligence models that learn to perform tasks by analyzing training examples. While these methods have shown to be able to perform on par with human doctors for complicated medical diagnosis tasks, such as skin cancer classification[1], they suffer from a fundamental drawback: The need for large amount of annotated samples required to develop the networks. FastRAI aims to bypass this, by leveraging the TUM’s expertise in simulating artificial X-rays from CT scans[2]. Thereby a large amount of simulated X-rays and corresponding labels can be generated using only a smaller amount of labelled CT scans. These simulated scans can then be used to train neural networks that are able to analyze real scans.[3]

The task of the networks is then to provide the following, clinically relevant information:

a) Provide a result with a high positive predictive value regarding COVID-19 infection from a standing chest X-ray (bilateral). Thus, patients could be stratified as “highly likely having Covid-19 infection” early on within minutes in order to receive the right isolation and treatment. This is clinically highly relevant since FastRAi will alleviate the need for Chest-CT or PCR availability. FastRAi therefore will be ideal for screening especially in countries with economic restraints.

b) Provide better insights from bed- chest X- rays. Chest X- rays taken of a bedridden patient are especially difficult since rays enter from the front. Problems here are mediocre quality due to obliteration of the „heart shadow“ and because of foreign materials (cables, tubes, etc) attached to the patient. Plus, a second plane (i.e. lateral image) of the patient is not possible, reducing the information about the pathologic findings. With 2 oblique images, usually not readable by the human eye, a model could derive better information about findings or offer an improved visualization to the care-giver, even comparing it to prior imaging data of the same patient (e.g. chest CT). This would provide a massive benefit (mobile X-ray has higher availability and is much less complicated to apply to severely sick patients).

[1] Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639), 115-118.

[2] Unberath, M., Zaech, J. N., Lee, S. C., Bier, B., Fotouhi, J., Armand, M., & Navab, N. (2018, September). DeepDRR–a catalyst for machine learning in fluoroscopy-guided procedures. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 98-106). Springer, Cham.

[3] Unberath, M., Zaech, J. N., Gao, C., Bier, B., Goldmann, F., Lee, S. C., … & Navab, N. (2019). Enabling machine learning in x-ray-based procedures via realistic simulation of image formation. International journal of computer assisted radiology and surgery, 14(9), 1517-1528.

The Consortium


deepc is a Munich-based technology startup with a team of AI researchers, medical professionals, developers, and business analysts that has its expertise in the creation of AI solutions in healthcare as well as their deployment into the clinical workflow. deepc provides the portal and infrastructure for radiologists to upload data and receive structured reports from the consortium’s radiologists. Our board-certified radiologists use deepc’s novel AI-augmented reporting software solution for the annotation and reporting of the uploaded imaging studies.

TU Munich

TU Munich’s chair of Computer Aided Medical Procedures and Augmented Reality is an expert department for AI in medical image processing. Their job is to create the final AI model that will deliver superior results and information from analyzing a simple thoracic X-ray.

M3i GmbH

M3i contributes its MxDB Digital Biobank and its unique services around the creation of medical AI training data to the consortium. This guarantees the processing of anonymized health data in a quality controlled, data protection and ethical compliant way. Furthermore, labels (such as segmentations of lesions etc.) are generated in an efficient and quality controlled process. The result are „gold standard“ training datasets that are used not only to train but also to validate and test the resulting model.

Moreover, M3i hires board certified radiologists to not only assure data quality but also to generate structured reports that are then provided to the radiologists in clinics that need second readings and expert help due to increased workloads.

Funded by