Artificial intelligence (AI) is being increasingly leveraged to enhance the analysis of medical imaging data, particularly in the realm of cancer diagnostics. A recent international competition, known as autoPET, has showcased this potential, with researchers from the Karlsruhe Institute of Technology (KIT) achieving a commendable fifth place among 27 competing teams. The findings from this competition were detailed in an article published in the journal Nature Machine Intelligence, highlighting the capabilities of algorithms in identifying tumor lesions through positron emission tomography (PET) and computed tomography (CT).
Medical imaging plays an indispensable role in cancer diagnosis. Accurately identifying the location, size, and type of tumors is critical when determining appropriate therapy options. PET and CT are key techniques in this process. PET enables the visualization of metabolic functions within the body using radionuclides. Specifically, malignant tumors, which display a significantly elevated metabolic activity compared to benign tissues, can be tracked using radioactively labelled glucose, typically fluorine-18-deoxyglucose (FDG). Meanwhile, CT scans layer the body image by image using X-ray technology to delineate anatomical structures and pinpoint tumor locations.
The analysis of these images, particularly in patients with numerous lesions, is traditionally labor-intensive. Doctors are required to manually mark tumour lesions on 2D images, a process that can be exceedingly time-consuming. Professor Rainer Stiefelhagen, the head of the Computer Vision for Human-Computer Interaction Lab at KIT, underscored the advantages of automation, stating, “Automated evaluation using an algorithm would save an enormous amount of time and improve the results.”
In the autoPET competition, which took place in 2022, Stiefelhagen, alongside doctoral student Zdravko Marinov and others from KIT and the IKIM Institute for Artificial Intelligence in Medicine in Essen, designed algorithms tasked with the automatic segmentation of metabolically active tumor lesions detected on whole-body PET/CT scans. The competition, organised by Tübingen University Hospital and LMU Hospital Munich, provided teams with access to a rich dataset of annotated PET/CT images, enabling them to train their algorithms. Each of the final submissions utilised deep learning techniques, which involve complex, multilayered neural networks to discern patterns from vast data sets.
The researchers reported that an ensemble of top-performing algorithms significantly outperformed individual models in detecting tumour lesions. Stiefelhagen explained that, “While the performance of the algorithms in image data evaluation partly depends indeed on the quantity and quality of the data, the algorithm design is another crucial factor, for example with regard to the decisions made in the post-processing of the predicted segmentation.”
The findings indicate that while considerable strides have been made in automating medical image analysis, further investigations are essential to optimise these algorithms, ensuring their robustness against external factors for effective clinical application. Researchers are aiming for a future where the analysis of PET and CT imaging data can be entirely automated.
The noteworthy achievements by the KIT research team reflect a growing trend towards integrating AI into medical practices, thus offering potential transformative benefits in the diagnosis and treatment of cancer. As advancements in deep learning and other related technologies progress, the implications for medical diagnostics and patient care are expected to expand significantly.
Source: Noah Wire Services