Visiting graduate researchers Louan Bardou and Emma Duprat presented at the recent UCSF Center for Intelligent Imaging (ci2) SRG Pillar Meeting. Both spent recent months as interns in the Rauschecker-Sugrue Lab in the UCSF Department of Radiology and Biomedical Imaging while completing their studies at École Polytechnique in France. Bardou presented his work on detecting and characterizing white matter abnormalities in adolescents, while Duprat presented her research on decoding segmentation uncertainty in H3K27M-mutant diffuse midline glioma.
White matter abnormalities are patches of white matter that look different from the surrounding tissue and can reflect injuries ranging from vascular problems to inflammation. Some are old, and some are still changing, and a single scan cannot reveal which is which, so understanding them means measuring their size, shape, and change over time across many cases. Bardou, who worked with Leo Sugrue, MD, PhD, and Pierre Nédelec, turned to the Adolescent Brain Cognitive Development (ABCD) Study, which follows more than 11,000 children between the ages of 9 and 17 with scans every two years. Radiologists flag a white matter abnormality in about 7% of sessions, but they do so manually and by eye, which leaves each lesion’s detailed characteristics buried in the imaging data.
Most tools for segmenting these lesions rely on FLAIR, an MRI sequence that makes lesions easier to see, and reach Dice scores of around 0.6. Because ABCD never acquired FLAIR, Bardou set out to match that performance without it. He hand-segmented more than 60 subjects, used deep-learning-assisted annotation in ITK-SNAP to label more, and reviewed each one manually. Training a separate model per sequence, he found that T2-weighted imaging performed best, reaching a Dice score of 0.55 that is comparable to FLAIR-based methods. The work produced more than 1,800 reviewed lesions and an automated pipeline, now available on GitHub, that other researchers can reuse. Bardou grouped the lesions by shape into round, elongated, punctate, and branching forms and used radiologists’ longitudinal reports to track which were new, growing, shrinking, or disappearing across five time points. He noted that the next step is prediction rather than observation, using diffusion imaging to help tell active lesions from settled ones.
Diffuse midline glioma is a rare and aggressive brain tumor that accounts for roughly 10% of pediatric tumors, mostly affects young children, and carries a median survival of only eight to 11 months. It grows in the midline of the brain, including the thalamus and pons, and its infiltrative growth gives it a heterogeneous MRI appearance that complicates treatment. Most patients carry the H3K27M mutation, which comes in two forms. The H3.3 subtype accounts for most cases and tends to produce more diffuse, infiltrative tumors with worse outcomes, while the H3.1 subtype tends to be more circumscribed and localized to the pons in younger children. Because researchers are exploring treatment that differ by subtype, knowing a patient’s subtype matters, yet it currently requires a biopsy of a tumor deep in the brain.
Duprat, who worked with Andreas Rauschecker, MD, PhD, and Pierre Nédelec, asked whether segmentation uncertainty could serve as a non-invasive marker of subtype. Working from the UCSF cohort of 260 patients, of whom 121 had radiologist segmentations and fewer than 100 had a known subtype, she built on an ensemble framework the lab had applied to meningiomas. Rather than producing a single mask, it runs several models in parallel and assigns each voxel a probability of belonging to the tumor, from which she computed epistemic uncertainty, the variance between models, and aleatoric uncertainty, based on the mean of the probability distributions. Epistemic uncertainty was significantly higher in H3.3 tumors and concentrated at the tumor border, consistent with that subtype’s infiltrative growth, and the pattern held after accounting for age, tumor volume, and location. Tests against survival, Ki-67 proliferation, and other mutations came back negative, leaving subtype as the one consistent association. Because radiomic features classified subtype about as accurately as the uncertainty measures, Duprat concluded that epistemic uncertainty encodes tumor morphology well enough to partly separate the two subtypes without any molecular label.
Visit the ci2 events page to learn more about upcoming SRG Pillar meetings.
