- [In German only] Peeken JC, Combs SE. Anwendung künstlicher Intelligenz in der Radioonkologie. Zielvolumendefinition und Organsegmentierung. Onkologie 2023; 29: 876–882.
The actual radiotherapy is preceded by complex planning with precise determination of the target volume and the neighbouring regions. The definition of the PTV is a central task of the radiotherapist. Attempts to support this process with image processing software have been around for a long time. However, it was the development of deep learning that brought decisive progress in this field. Deep learning is set by the architecture of several interconnected layers forms a kind of artificial neural network that can process enormous amounts of data and "learn" from it.
Radio-oncologists Peeks and Combs have looked into the extent to which AI can already help with radiotherapy planning.
The first approved software solutions for automated contouring and segmentation of anatomical structures already exist. They offer clear advantages in clinical application, as numerous studies have now shown. For example, the use of auto-contouring tools can save a significant amount of time. At the same time, fewer manual corrections are necessary. In addition, clinical processes can be standardised and therefore improved.
The fact that AI in radiation oncology is not just a theoretical benefit, but can ultimately also improve the outcome, was shown in a study on the radiotherapy of lung carcinomas. Automated cardiac structure contouring reduced the average radiation exposure compared to manual acquisition. This in turn correlated with longer overall survival.
A high segmentation quality was also achieved with neural networks when contouring brain metastases and accompanying oedema. And even clinical target volume (CTV) of the regional lymphatic drainage channels, which are often not clearly delineated and comprise different anatomical structures, have already been successfully segmented using AI.
Despite all the technical progress, manual control by the radiotherapist remains indispensable, according to the authors. On the one hand, software solutions are still reaching their limits, especially for small structures and organs with variable positions. But the overall radiation planning process is so complex, with various imaging methods and numerous clinical-pathological influences, that it cannot (yet) be fully automated.
And even if it should one day be possible to integrate all imaging methods and clinical-pathological influences into the software, the radiotherapist will ultimately have to check the result and correct it manually if necessary.
AI applications in the field of radiation oncology are developing rapidly. Approved software solutions can already support treatment planning. In future, radiotherapy could be tailored even more individually to patients with the help of AI. But one thing is clear: the final say lies with the radiotherapists.