De Sutter, S. (2025). Segmentation of glioblastoma from multi-modal medical imaging: Towards revealing tumor infiltration. [PhD Thesis, Vrije Universiteit Brussel]. Crazy Copy Center Productions.

@phdthesis{9a74e654f95240c08998738e89cd2d28,
title = "Segmentation of glioblastoma from multi-modal medical imaging: Towards revealing tumor infiltration",
abstract = "Glioblastoma is the most common and most aggressive form of primary brain tumors. As currently margins for maximal safe resection and radiotherapy planning are based on a contrast-enhancing (CE) lesion defined on magnetic resonance imaging (MRI), tumor cells that infiltrate the healthy tissue beyond the CE component are not targeted during treatment and become a source of tumor recurrence, of which the majority of patients ultimately fall victim. Despite this understanding, the detection of infiltration on medical imaging remains elusive. This work aimed to leverage the power of artificial intelligence to reveal complex data patterns for achieving a better segmentation of glioblastoma and its infiltration from multi-modal medical imaging. Initially, an intuitive, probabilistic segmentation approach was explored to challenge the conventional and omnipresent deterministic segmentation of glioma on MRI, of which the binary labels do not properly reflect the underlying biology of a tumor with such a diffuse growth. This brain tissue mapping model, based on conventional MRI sequences, including T1-weighted (T1), contrast-enhanced T1-weighted (T1CE), T2-weighted (T2) and T2-fluid-attenuated inversion recovery (FLAIR), was evaluated in terms of segmentation accuracy of the tumor and its subregions, and in the visualization of possible infiltration. While the model achieved good accuracy for detection of the whole tumor, lower accuracy was found for tumor subregions compared to state-of-the-art deep learning (DL) models. Visual inspection of tumor probability maps revealed high probability values within the CE lesion, with lower values extending into the surrounding edema region. This pattern aligns with our hypothesis that these probability maps potentially reflect cell density and can model gradual tissue transitions. However, interpretation remained ambiguous due to the limited segmentation accuracy and the lack of validation possibilities. Subsequently, we explored DL models for deterministic segmentation tasks. While state-of-the-art automated glioblastoma segmentation is tradition- ally performed on a four-modality input, we postulate that information redundancy is present within combinations of these images, possibly redu- cing the performance of these models. In addition, the risk of encountering missing data rises as the number of required input modalities increases. Therefore, through the evaluation of segmentation accuracy and epistemic uncertainty of multiple segmentation models, differing only in their amount and combinations of input modalities, the relevance of each of the modalities concerning glioblastoma segmentation was brought to light. Results showed that T1CE and FLAIR were sufficient to reach accuracies comparable to the four-modality model, and can serve as a minimal-input alternative to the full-input configuration. While additional modalities beyond this did not improve – and even deteriorated – accuracy, their presence was found to reduce segmentation uncertainty. Although according to multiple biopsy-controlled studies, positron emission tomography (PET) with amino acid tracer O-(2-[18F]fluoroethyl)-L-tyrosine ([18F]FET) reportedly allows better estimation of the tumor extension com- pared to the CE boundaries found on MRI, the automated segmentation of the lesion on this imaging modality is ill studied. Therefore, we explored the use of deep learning for a robust and automated method for glioblastoma segmentation from this imaging modality. While results comparable to the current state of the art were obtained, our results indicate that the lack of reproducible ground truth restricts the maximum achievable accuracy for automated glioblastoma segmentation on [18F]FET PET for all networks. The previous works lay the foundation for the integration of information from both MRI and [18F]FET PET, that allow a more comprehensive characteriz- ation of the tumor{\textquoteright}s composition and an estimation of the infiltrative region. To demonstrate this vision, we explored simultaneous segmentation of labels defined on MRI and [18F]FET PET, with the PET-positive lesion beyond the CE region functioning as a surrogate for infiltration labeling. Moreover, we investigated the possibility of predicting such label from MRI alone, allowing better definition of the tumor{\textquoteright}s extent while eliminating the need for PET acquisition. Although results indicate the investigated approaches allow to partially compensate for the absence of PET information, prediction based on solely on MRI does not achieve the needed accuracy, supporting the inclusion of the acquisition of [18F]FET PET for glioblastoma management. In conclusion, this study highlights the potential of artificial intelligence to enhance glioblastoma imaging analysis and explores tools that build towards improved detection of infiltration. By examining probabilistic and deep learning models, we identified critical MRI modalities and explored [18F]FET PET{\textquoteright}s ability to better delineate a tumor{\textquoteright}s extent. Our findings set the stage for a combined MRI-PET approach, aiming to improve tumor characterization and guide more effective clinical strategies.",
author = "\{De Sutter\}, Selene",
year = "2025",
language = "English",
isbn = "9789464948851",
publisher = "Crazy Copy Center Productions",
address = "Belgium",
school = "Vrije Universiteit Brussel",
}