This project aims to develop a multiscale atlas of the brain by integrating 2D and 3D multimodal data from both normal and pathological samples. Using datasets such as the Allen Brain Atlas, digital pathology, high-resolution microscopy, and brain MRI, we will explore the correlation between cellular, spatial, and anatomical features across different scales. Novel methods will be developed for 2D to 3D registration, focusing on aligning histological images with MRI data to localize cellular details within a 3D anatomical context.
The atlas will be applied to brain tumor datasets, including TCGA, EBRAINS, and private collections, to capture complex spatial and molecular relationships. By integrating advanced graph-based models, this project will enhance understanding of both normal brain structure and tumor heterogeneity, with potential applications in personalized treatments for neuro-oncological diseases.

The proposed PhD project aims to develop a comprehensive multiscale atlas of both normal and pathological brains through the integration of multimodal datasets. The primary goal is to establish a unified framework that accurately represents anatomical and molecular correlations across different spatial scales, from cellular features to volumetric anatomical structures. This atlas will provide a foundational tool for understanding the complex organization of the brain, bridging cellular-level insights with broader anatomical features. The project will initially focus on studying normal brain anatomy using high-resolution datasets such as the Allen Brain Atlas, which contains detailed molecular, anatomical, and spatial data across multiple species. The Allen Brain Atlas includes high-resolution in situ hybridization (ISH) data, gene expression patterns, and corresponding anatomical mappings, making it an ideal starting point for understanding the baseline organization of the brain.
The project will then expand into pathological conditions, particularly focusing on brain tumors. We will incorporate large-scale, publicly available datasets such as The Cancer Genome Atlas (TCGA), which includes multi-omics data (genomics, transcriptomics, epigenomics) alongside histopathological and MRI images from thousands of tumor samples. Additionally, the EBRAINS platform will provide access to various imaging datasets that offer a macroscopic view of brain anatomy and function, including MRI data. The project will also leverage private collections from collaborating institutions, encompassing over 5,000 patient samples, providing both multi-modal imaging and associated clinical metadata, such as survival outcomes, treatment regimens, and molecular profiles. These resources will allow us to create a multiscale model that can be applied to the study of tumor heterogeneity, disease progression, and response to therapy.
The challenge of integrating such diverse modalities lies in their varying spatial resolutions and data structures. For instance, digital pathology images captured at the cellular level (2D) operate at a vastly different resolution than MRI scans (3D) that provide a volumetric view of the entire brain. Additionally, the dimensionality of these datasets varies significantly, with molecular profiles containing thousands of features per sample. Therefore, one of the key objectives of this project is to develop novel methods for 2D to 3D registration, which is critical for aligning the various modalities and integrating them into a coherent multiscale atlas. This will involve aligning high-resolution microscopic images from digital pathology and high-resolution microscopy with volumetric MRI data, which captures macroscopic brain structures.
We will also develop methods for reconstructing 3D representations from stacked 2D Whole Slide Images (WSIs), which provide rich details at the cellular and subcellular levels. By stacking these WSIs, we can create a volumetric representation that bridges the gap between cellular and anatomical scales. These innovations will enable us to spatially localize cellular features—such as specific cell types, gene expression patterns, and tumor microenvironments—within a 3D anatomical context, providing a more detailed and comprehensive understanding of both normal brain organization and pathological changes that occur in diseases like glioblastoma.

The methodological foundation of this project will focus on graph-based models for understanding spatial relationships between cells and structures in both physiological and tumor contexts. Specifically, we will apply Graph Neural Networks (GNNs) to encode spatial correlations between cells within the 3D brain atlas. This will enable us to capture complex spatial interactions that are difficult to detect using conventional approaches. The graph-based representations will support understanding how cellular architecture and spatial patterns change in disease states, such as tumor heterogeneity.
A key aspect of the project will also involve the use of Graph Attention Networks (GATs) for self-supervised learning, particularly in the context of multimodal data registration. GATs will be leveraged to integrate different data types during the 2D to 3D registration process, specifically to learn and preserve important spatial relationships between 2D histological images and their corresponding 3D MRI volumes. These networks will help refine the registration process by improving feature matching between modalities, facilitating a more accurate integration of cellular and structural information.
In addition to registration, another major goal of the project is the clinical validation of the developed atlas. This will involve cross-referencing the atlas with real-world clinical outcomes and molecular data, specifically focusing on brain tumor datasets. By analyzing spatial and molecular data within this multiscale framework, we aim to identify important correlations between tissue morphology and molecular phenotypes, such as gene expression profiles and mutations. The use of high-dimensional tabular data will be incorporated to provide a holistic view of how molecular changes at the cellular level are reflected in tissue morphology. Techniques such as saliency mapping and feature attribution will aid in the explainability of these models, highlighting key features that correlate with clinical outcomes.
The validation of the atlas across normal and pathological samples will allow us to assess the robustness and applicability of our models. By incorporating molecular data, we will aim to refine our models to better predict patient outcomes and treatment responses. This project will not only contribute to neuro-oncology research but also provide a framework applicable to broader neurological diseases and other complex biological systems.

PhD Scope Summary:
1. Data Gathering and Preprocessing: Gather and preprocess multimodal datasets from publicly available sources (e.g., Allen Brain Atlas, TCGA, EBRAINS), ensuring consistency and preparing them for further analysis.
2. 2D-3D Registration: Develop novel methods for aligning 2D histopathological images with 3D MRI data, focusing on multimodal registration and feature preservation.
3. Graph-Based Self-Supervised Models: Implement GNNs and GATs for spatial correlation modeling and use them to guide 2D-3D registration as well as multimodal integration.
4. Clinical Validation: Validate the generated atlas by correlating its predictions with clinical outcomes and molecular data, particularly focusing on brain tumors.