Artificial intelligence: Algorithms improve medical image analysis

Imaging technology plays a key role in cancer diagnosis. Accurately determining the location, size, and type of tumor is critical to selecting the correct treatment. The most important imaging techniques include positron tomography (PET) and computed tomography (CT). PET uses radioactive nuclei to visualize metabolic processes in the body. The metabolic rate of malignant tumors is significantly higher than that of benign tissues. Radiolabeled glucose, usually fluoro-18-deoxyglucose (FDG), is used for this purpose. In CT, the body is scanned layer by layer in an X-ray tube to visualize the anatomy and locate tumors.

Automation saves time and improves assessments

Cancer patients sometimes have hundreds of lesions, pathological changes caused by tumor growth. In order to obtain a uniform image, it is necessary to capture all lesions. Doctors determine the size of tumor lesions by manually labeling 2D slice images, which is an extremely time-consuming task. “Automated evaluation using algorithms will save a lot of time and improve results,” explains Professor Rainer Stiefelhagen, Head of Computer Vision (cv:hci) at the Human-Computer Interaction Laboratory at Karlsruhe Institute of Technology.

Rainer Stiefelhagen and cv:hci doctoral student Zdravko Marinov participated in the 2022 International AutoPET Competition and took fifth place among 27 teams with 359 participants from around the world. The researchers from Karlsruhe formed a team with Professors Jens Kleesiek and Lars Heiliger from IKIM (Institute for Artificial Intelligence in Medicine) Essen. AutoPET, organized by the University Hospital Tübingen and the University Hospital LMU in Munich, combines imaging and machine learning. The task is to automatically segment metabolically active tumor lesions displayed on whole-body PET/CT. For algorithm training, participating teams have access to large annotated PET/CT datasets. All algorithms submitted in the final stage are based on deep learning methods. This is a variation of machine learning that uses layers of artificial neural networks to identify complex patterns and correlations in large amounts of data. The seven best teams in the autoPET competition have now reported on possibilities in automated analysis of medical imaging data natural machine intelligence Magazine.

Algorithm integration performs well in detecting tumor lesions

As the researchers explain in their publication, it turns out that ensembles of top algorithms outperform a single algorithm. The algorithm set can effectively and accurately detect tumor lesions. “While it is true that the performance of algorithms in image data evaluation depends in part on the quantity and quality of the data, algorithm design is another critical factor, such as decisions made in post-processing for predictive segmentation,” said Stiffelhagen explained. Further research is needed to improve the algorithms and make them more resistant to external influences so that they can be used in daily clinical practice. The goal is to fully automate the analysis of medical PET and CT image data in the near future.