UNet-verstärkte, Deep Learning-basierte Vorhersage der intraoperativen 5-ALA-Fluoreszenz mittels präoperativer multimodaler MRT bei niedriggradigen Gliomen.

Only 20-30% of lower-grade gliomas typically have fluorescence after 5-aminolevulinic acid (5-ALA) administration. This can be increased by doubling the administered dose of 5-ALA. We aimed to analyze if a deep learning model can predict intraoperative fluorescence based on preoperative multimodal magnetic resonance imaging (MRI). Accurately identifying tumor regions during surgery is crucial for identifying regions of anaplastic foci, as they tend to fluoresce, allowing surgeons to precisely deliver malignant tissue for histopathological evaluation and avoid undergrading.

The MRI images consisted of T1, T1-post gadolinium, and FLAIR. The images were standardized through a preprocessing pipeline involving isotropic transformation, bias field correction, registration to T1-post gadolinium space, alignment to a brain atlas, and skull stripping using ANTs and FSL software packages. The preprocessed MRIs were fed into the UNet, which was initially developed for tumor segmentation for each subject. We used the outputs of the bottom layer of the UNet in the Variational Autoencoder as features for classification. Using the partial least square discriminant analysis algorithm, we identified and utilized the most effective features in a Random Forest classifier. We allocated ~80% of the data for training purposes.

We included a cohort of 170 patients categorized as positive (n=89) or negative (n=81). The performance of our proposed approach is evaluated using key metrics. The optimal results were obtained by employing 18 top-performing features, resulting in an accuracy of 79% with the following associated confusion matrix:

[14 3

4 13]

Our findings highlight the potential of a UNet model, coupled with a random forest classifier, for intraoperative fluorescence prediction. We achieved a good accuracy by using advanced techniques such as DL-based tumor segmentation and Variational Autoencoder for radiomics feature extraction. While the model can still be improved, it has the potential for evaluating when to administer 5-ALA to tumors lacking typical imaging features of high-grade gliomas.