Magnetic resonance imaging (MRI) has emerged as a valuable non-invasive tool to characterize the large-scale architecture of the brain. Structural MRI provides insights into the brain’s morphology, while diffusion weighted imaging (DWI) allows to track the white matter pathways in-vivo (Wang and He, 2024, Le Bihan, 2003).

Of note, single white matter pathways do not exist in isolation, but group into coherent formations known as bundles. Their identification is of pivotal importance, as white matter bundles (or tracts) support the characterization of healthy functioning as well as the derangement occurring in neurological disorders (Yamada et al., 2009, Kamagata et al., 2021, Ghazi et al., 2023). For brain tumor patients, the segmentation of white matter bundles – also known as virtual dissection – is particularly important, as it provides neurosurgeons with a neuro-biologically grounded reference driving the selection of brain areas not to be damaged during surgery (Yamada et al., 2009, Ghazi et al., 2023). As such, virtual dissection plays a crucial role in avoiding surgery-induced permanent cognitive impairments (Yang et al., 2021).

White matter bundle segmentation is typically performed manually using the patient’s specific MRI acquisition, combining the identification of anatomical landmarks and the manual selection of white fibers reconstructed via DWI and fiber tracking (Andica et al., 2023, Wakana et al., 2007). In the neurosurgical theater, the virtually dissected bundle is used as an anatomical reference driving tumor resection during surgery (Cho et al., 2014). Given the requested amount of knowledge about human brain anatomy and the dramatic consequences arising from an erroneous segmentation, the pre-surgical identification of clinically relevant white matter tracts is the responsibility of dedicated and ad-hoc trained personnel. Nonetheless, manual bundle segmentation is time-consuming, prone to errors, and highly operator dependent (Tunç et al., 2016, Joshi et al., 2024).

To address these critical issues, the research community has produced a significant effort to develop automated and reliable methods for the bundle dissection (Yendiki et al., 2011, Ghazi et al., 2023).

Automated methods for white matter bundle segmentation can be classified into two main strategies: atlas-based and learning-based. The former exploits manually curated and normative references delineating prototypical bundle representations (Garyfallidis et al., 2018, Zhang et al., 2018). Learning-based models, on the other hand, use a collection of ground-truth examples to automatically extract the set of rules governing the segmentation process (Wasserthal et al., 2018, Bertò et al., 2021). In the context of glioma research, atlas-based methods are a promising alternative when bundles are not heavily deformed by tumors. In cases of large tumor-induced anatomical alterations, atlas-based methods developed on healthy individuals perform poorly, as they cannot account – by definition – for large anatomical alterations. Analogously, the development of learning-based methods for white matter tract segmentation in glioma patients is hindered by the lack of dedicated large-scale and publicly available clinical datasets. Taken together, a reliable solution to automatically extract clinically relevant white matter bundles in brain tumor patients is currently missing but it is an urgent clinical need.

A conceptually simple strategy to address this gap is learning a model on a reference population of healthy individuals, and then adapting to a custom (and usually much smaller) sample of interest. This strategy – known as transfer learning – yielded promising results across several clinical neuro-imaging domains, as it allows to compensate for a difference in data distribution across samples — the so-called domain shift (Ismaila et al., 2022). However, transfer learning has never been investigated in the context of white matter bundle segmentation in glioma patients.

Leveraging the unprecedented availability of manually curated tract segmentations in glioma patients, in the present work we investigated the efficacy of transfer learning in the context of white matter bundle segmentation in this clinical sample. Specifically, we first trained a convolution neural network on a large-scale dataset encompassing hundreds of healthy individuals and several manually annotated white matter bundles. We then applied transfer learning to two distinct samples of glioma patients. The main contributions of our work are twofold: (i) We developed a novel metric enabling the distinction and quantification of domain shift occurring due to the tumor (biological bias) from systematic variations (different MRI scanners, acquisition parameters, and pre-processing, systematic bias) with single subject resolution. (ii) Importantly, we show that transfer learning could recover the domain shift related to systematic variations, but not the spatial re-arrangement of the bundles induced by the tumor.

Our findings were coherent across the five white matter bundles and three different input modalities tested, highlighting their robustness and generalizability.

Taken together, our results advance the state-of-the-art by showing that white matter bundle segmentation in glioma patients is a challenging task that cannot be addressed with simple transfer learning approaches.