Algorithms can detect tumors

Algorithms can help localize tumours. This is the result of a study that evaluated 27 different algorithms. These were developed as part of the International Machine Learning Challenge (autoPET), an international initiative organized by the University Hospital of Tübingen and the University Hospital of Munich. Until now, radiologists and nuclear medicine specialists have manually recorded the size of so-called tumor lesions in 2D slice images. This process is very time-consuming. The aim of the challenge was to automate the process. The results of the study show that algorithms have the potential to be used for this task in everyday clinical practice in the future.

Knowing the exact size, type and volume of tumors is important for recommending the right treatment to cancer patients. This is based on images taken using positron emission tomography (PET) and computer tomography (CT). CT visualizes the anatomy of the body, while PET measures the metabolic activity of tissue. Until now, radiologists and nuclear medicine specialists have manually marked the contours of individual tumors in 2D slice images in order to determine their volume. In some cases, cancer patients have several hundred lesions - pathological changes caused by the growth of a tumor. In order to obtain a uniform picture, all lesions would have to be recorded. "Doctors would then need several hours to mark the lesions. With many thousands of cancer patients per year at Tübingen University Hospital, this is very time-consuming work," explains Prof. Dr. Thomas Küstner, head of the Medical Image and Data Analysis Lab research group in the Department of Diagnostic and Interventional Radiology at Tübingen University Hospital. "It would be better if an algorithm could take over the analysis automatically."

In 2021, Küstner and his colleague Prof. Dr. Sergios Gatidis approached Prof. Dr. Clemens Cyran and Prof. Dr. Michael Ingrisch from LMU Klinikum München. The idea: to develop a challenge that combines imaging with machine learning. The first autoPET Challenge was organized in 2022 and ran from April to September 2022. Researchers from all over the world were invited to develop algorithms for the evaluation of tumor lesions. 359 participants in 27 teams submitted their solutions. The results were presented at the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in September 2022. "We were able to reach hundreds of scientists worldwide and motivate them to work on this important issue. This was only possible as part of a challenge," explains Gatidis. All of the submitted algorithms use deep learning, a form of artificial intelligence, to identify the contours of the lesions. Neural networks are used to recognize complex patterns and correlations in large amounts of data. Küstner and his team subsequently identified the most promising algorithms.

Aiming for complete automation

The evaluation of the algorithms shows that the algorithms can detect tumor lesions precisely and efficiently. "The challenge was an important first step in identifying promising algorithms in this field," emphasizes Ingrisch. However, the performance of the algorithms depends heavily on the quality of the PET and CT images. The better the algorithms can handle different data sets, the more robust the algorithms are against external influencing factors such as variations in the images, the imaging scanners or the radioactive markers. "Our goal is to fully automate the analysis of CT and PET imaging in the next few years," explains Cyran. Further work is therefore necessary to improve the algorithms and make them more robust.

"We are delighted that, after the initial pioneering work, we can now also contribute scientifically to the continuation of this successful project. In the meantime, we have been able to set up a motivated joint working group of radiology and nuclear medicine clinics with regular exchanges," says Matthias Brendel, acting director of the Clinic and Polyclinic for Nuclear Medicine at LMU University Hospital. The team from Tübingen and Munich will evaluate the results again and take further steps to be able to use the algorithms in everyday clinical practice in the near future.

Original publication:
Results from the autoPET challenge on fully automated lesion segmentation in oncologic PET/CT imaging
Sergios Gatidis, Marcel Früh, Matthias P. Fabritius, Sijing Gu, Konstantin Nikolaou, Christian la Fougère, Jin Ye, Junjun He, Yige Peng, Lei Bi, Jun Ma, Bo Wang, Jia Zhang, Yukun Huang, Lars Heiliger, Zdravko Marinov, Rainer Stiefelhagen, Jan Egger, Jens Kleesiek, Ludovic Sibille, Lei Xiang, Simone Bendazzoli, Mehdi Astaraki, Michael Ingrisch, Clemens C. Cyran, Thomas Küstner

Nature Machine Intelligence
DOI: 10.1038/s42256-024-00912-9