A study led by researchers from the BDSLab group at the ITACA Institute of the Universitat Politècnica de València highlights the need to adapt artificial intelligence technologies to the particularities of each medical institution to maximise their impact on patient care, especially in the treatment and segmentation of glioblastomas, one of the most aggressive brain tumours.
This is the main conclusion of an article published in the International Journal of Medical Informatics. The research has had the participation of F. Javier Gil-Terrón, Pablo Ferri, Víctor Montosa, María Gómez and Carles López, led by Juan M. García and Elíes Fuster, as well as Pau Martí, from the Universitat de les Illes Balears.
Importance of AI in the treatment of glioblastomas
The study is based on the reality that Magnetic resonance imaging (MRI) has revolutionised the diagnosis and treatment of brain tumours, allowing non-invasive access to crucial information for planning surgical interventions and monitoring treatment progress, especially in cases of glioblastoma. However, manual segmentation of images to accurately identify affected areas can be complex due to the three-dimensional and multi-parametric nature of the images.
To improve this process, advances in deep learning, especially in convolutional neural networks (CNNs), have facilitated the development of automatic models that provide fast and accurate segmentations of brain tumours, optimising the treatment of patients with glioblastoma, a particularly lethal type of tumour.
In this context, the work analysed how data variability between different medical centres can affect the accuracy of learning models in glioblastoma segmentation, due to the phenomenon known as dataset shift. Specifically, they addressed variability in the differences in magnetic resonance imaging (MRI) images obtained at different centres, in the segmentation criteria applied by experts, and in the tumour composition itself.
In contrast to generalist models, which are trained on data from multiple sources and may lose effectiveness when confronted with a specific clinical setting, personalised models address the unique characteristics of each centre. This is essential in brain tumour segmentation, where diagnostic accuracy can directly influence therapeutic decisions,’ says F. Javier Gil-Terrón, lead author of the study.
Work done in the research
‘Adapting the models not only optimises performance, but also allows healthcare facilities to benefit from more reliable tools. By customising the models for each institution, variations in image acquisition practices, clinical protocols and segmentation criteria can be addressed, aspects that are often overlooked in generalist approaches,’ adds F. Javier Gil-Terrón.
The research was conducted using the BraTS 2021 database, which includes more than 1,200 cases from 23 medical centres. From this data, 155 learning models based on convolutional neural networks were developed and compared to evaluate the performance of different configurations.
For all these reasons, the authors of the study emphasise that the proposed approach supports the idea that artificial intelligence should be perceived as a flexible and adaptable tool, tuned to the specific requirements of each clinical setting.
‘Adapting the models to the reality of each medical centre not only improves accuracy, but also facilitates their implementation and acceptance by healthcare professionals,’ conclude the ITACA researchers.
Reference
Javier Gil-Terrón, Pablo Ferri, Víctor Montosa-i-Micó, María Gómez Mahiques, Carles Lopez-Mateu, Pau Martí, Juan M. García-Gómez, Elies Fuster-Garcia, Exploring the Trade-Off between generalist and specialized Models: A center-based comparative analysis for glioblastoma segmentation, International Journal of Medical Informatics, Volume 191, 2024, 105604, ISSN 1386-5056, https://doi.org/10.1016/j.ijmedinf.2024.105604.