Fraunhofer Institute for Digital Medicine MEVISPress Release
Fig. 1 A deep learning approach allows CT images to be compared automatically in the context of cancer treatment — so that even small changes to lung tumors can be detected and located with greater speed and precision.
Regular follow-up monitoring is crucial when it comes to treating cancer successfully. In the SPIRABENE project, Fraunhofer Institute for Digital Medicine MEVIS developed a deep learning-based software program which makes it possible to identify disease- and treatment-related changes in tumors in CT images with even greater speed and accuracy, thus improving the chances of recovery and making day-to-day clinical practice easier.
More and more thoracic CT scans are being performed worldwide in order to detect lung diseases such as bronchial carcinoma and tumor metastases at an early stage and allow them to be monitored. In Germany as well, they are now one of the most common radiological examinations: 800,000 lung CT scans were done in 2009, compared to 1.3 million in 2020.
The scans make it possible to identify even the most minor treatment effects and side effects, allowing the treatment to be optimized as a result. However, comparing the scans is a highly complex and time-consuming task which is susceptible to errors since radiologists often have to work under extreme time pressure when evaluating the images. Automatically establishing anatomical correspondence between the scans – a process known as registration – makes this easier.
Reliable comparisons of CT scans using deep learning
To optimize diagnostics and make day-to-day clinical practice easier for doctors, the SPIRABENE project – which involves collaboration between Fraunhofer MEVIS and its project partners jung diagnostics GmbH and University Medical Center Mainz UM and is funded by the German Federal Ministry of Education and Research (BMBF) – is focusing on artificial intelligence: “We have developed a deep learning-based software program which makes it possible to locate and measure lung lesions with greater precision and in a very short time, as well as allowing us to detect potential new lesions,” explains Jan Moltz, Key Research Engineer in Medical Image Analysis at Fraunhofer MEVIS.
To register the lungs, previous CT scans are compared with the latest images in order to establish anatomical correspondence. A particular challenge here is that two images of the same person look similar but are not identical – for example, due to variations in breathing when the scan was taken or possible weight loss due to the treatment.
Follow-up image monitoring is already supported by automated registration, but the use of deep learning makes it possible to compare the scans automatically with even greater speed and precision. The researchers optimized neural networks for this purpose. Moltz explains: “Our results show that 11 percent more tumors are automatically detected in the follow-up image with the help of AI compared to conventional software-based registration – and at less than one second, it takes a tenth of the time.” This also means that less computing power is required, which saves energy.
The researchers designed the fully automated image processing technology in collaboration with doctors from UM, enabling it to be integrated and used directly in existing clinical infrastructures. The software has already been tested and evaluated in day-to-day clinical practice and could soon be used in real-life applications. The long-term plan is for AI-supported follow-up monitoring to be used for the entire body.
Following the test run, Moltz is pleased with the initial success: “I feel motivated by the idea of working on software that is actually used in clinical practice and has a positive impact on doctors’ day-to-day work. The software also helps us to identify ineffective treatments promptly, avoid unnecessary side effects and costs, and increase the chances of recovery for patients.”
Press Release
- Press Release Fraunhofer Research News (fraunhofer.de)
- Our Solution: OncoChange (mevis.fraunhofer.de)