Boston University
Knowledge-guided representation learning for multimodal neuroimaging analysis
Abstract
dc:description.abstractArtificial intelligence (AI) frameworks based on deep learning have demonstrated impressive performance on curated neuroimaging benchmarks, especially magnetic resonance imaging (MRI), but translation to routine clinical practice remains limited by three persistent realities of real-world MRI: (i) distributional shift across institutions, scanners and acquisition protocols; (ii) modality incompleteness, where one or more MRI sequences are absent due to time, cost, motion or clinical triage; and (iii) the need for clinically aligned reasoning, where models must not only predict labels but also communicate findings in the structured language used by radiologists. This dissertation advances robust and clinically applicable neuroimaging AI by integrating structured domain knowledge across multiple stages of the modeling pipeline. First, we introduce a disease-driven domain generalization framework for neuro\-imaging-based assessment of Alzheimer's disease that encourages models to rely on biologically meaningful regions rather than cohort-specific shortcuts, improving performance under distribution shift. Second, we present an anatomy-guided augmentation strategy for brain tumor segmentation that leverages structural anatomical context to enhance robustness across diverse input configurations, achieving strong segmentation performance under challenging missing-modality conditions. Third, we develop a vision-language model for multi-sequence brain MRI interpretation across diverse neurological conditions, trained using large-scale report-based supervision and evaluated on multiple external clinical and research cohorts. We further characterize and mitigate modality hallucinations, i.e., unsupported sequence-specific claims during report generation, via a sentence-level detection framework and a training-free modality-aware faithful decoding strategy. Collectively, these contributions demonstrate that integrating pathology-, anatomy- and language-based domain knowledge into model design can improve generalization, resilience to missing data, and the clinical faithfulness of model outputs, providing a pathway toward more reliable neuroimaging AI in heterogeneous settings.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lteif, Diala
- Advisor dc:contributor.advisor
-
- Kolachalama, Vijaya B.
- ORCID dc:identifier.orcid
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53713