2021/10/01 | Research | Artificial Intelligence
Artificial Intelligence is used more and more extensively in image interpretation for diagnosis and treatment planning. A broad-based research team from Inselspital, Bern University Hospital and the University of Bern was able to demonstrate in a study that the current methods of qualifying AI for brain segmentation could be enhanced. Deviations as measured by means of currently used parameters do not correlate with clinically relevant changes of the radiation dose distribution. For widely supported implementation of AI based software there needs to be more focus on clinically relevant outcomes for it to provide real added value in terms of treatment quality.
The study focused on localizing and detecting the volume of tissues surrounding the brain tumor to be irradiated, the so-called organs at risk, the protection of which is given top priority in radiotherapy. Traditionally, CT and MRI images are interpreted by specially trained medical professionals who manually delineate the contours of organs at risk. This procedure is not only extremely time consuming, but also requires several physicians to analyze an image in order to exclude individual deviations or misinterpretations. In modern systems, this work is increasingly performed by automated software that interprets images based on the use of AI to plot the desired contours or calculate volumes. The present study investigated to which extent currently used metrics for automated image segmentation lead to clinically meaningful conclusions about dose distribution in organs at risk near brain tumors. Data from glioblastomas, a common, aggressive type of brain tumor, were used.
Surprisingly, the current methods used to determine the quality of AI based automated contouring do not have any predictive value on the quality of the treatment. Additionally, it was found that in many cases, mis-interpretations of the organs at risk have very little impact on the dose that would be delivered during radiotherapy. This means that the metrics used today in AI-assisted image evaluation could be reconsidered, and this changes the direction of improving AI methods and the implementation in day-to-day practice.
In a first step, the research group at the ARTORG Center for Biomedical Engineering Research led by Prof. Mauricio Reyes and the Department of Radiation Oncology led by PD Dr. med. Evelyn Herrmann selected viable cases of glioblastoma patients from the treatment history. Three specialists delineated reference organs at risk to serve as the basis for this study. Additionally multiple variations on the delineations of these organs were collected from multiple sources, including AI methods. Differences between the variations image interpretation and the expert findings were then determined by means of 20 currently available measures. In addition, it was assessed to what extent these differences had an effect on the dose delivered to these organs. The last step was to find the correlation between the 20 measures and the effect on the dose.
The study demonstrated the importance of systematic, interdisciplinary collaboration between AI and clinical experts. First author Robert Poel notes: ““The bases currently used for qualifying automatic image segmentation were tested in our study to determine whether they provide useful information on treatment quality. We found that we need to look for new measures that could exploit not only the acceleration of implementation but also the potential gain in quality from using AI.” Close, early collaboration between disciplines can bring further improvements here. Facilitating this collaboration is the aim of the newly founded Center for Artificial Intelligence in Medicine in Bern.
The use of AI-image interpretation in diagnosis and treatment makes sense and will undoubtedly become widespread. Prof. Dr. med. Daniel Aebersold, Chief Physician at the Department of Radiation Oncology and Head of the UCI University Cancer Center Inselspital describes a clear roadmap towards this goal: “The relevant disciplines must work closely together and at an early stage. The conditions for this are optimal on the Insel Campus. The close network of teams and locations that exists between the University, ARTORG and the hospital departments must be used to further align automated image segmentation with clinical issues. The recently published research project is an excellent example of this way of working.”
Link to the study
News article Inselspital, 1 October 2021
Medical Image Analysis