Brain cancer is a significant public health issue, affecting about 700,000 people in the U.S. and leading to over 80,000 new diagnoses each year. Despite advances in neuroimaging, genetics, and therapies, treatments remain largely ineffective due to tumor heterogeneity and delayed intervention. There is an urgent need for a shift toward personalized, precision, and predictive medicine (3P-medicine). Digital twins could revolutionize this approach, but current models lack effective neuroimaging biomarkers, longitudinal data, and computational tools for tumor detection and prognosis. This project aims to develop advanced frameworks for tumor twin modeling by: a) creating a comprehensive longitudinal brain tumor data cohort; b) developing a learning-based generative model for tumor progression using biological priors and multimodal imaging; and c) prototyping a digital twin of glioma to facilitate personalized radiation therapy planning.
The primary goal of this proposal is to leverage the exciting progress in brain imaging and machine learning to develop AI-powered computational tools as foundations for digital twin of brain tumor progression.
Aim 1 (Longitudinal brain tumor data collection) We will use both public datasets from GREI repository and our internal experimental dataset. The public databases feature over 1,600 longitudinal MRI data, providing rich spatiotemporal distributions of brain tumors, along with clinical, genetic, and pathological information. Specifically, Brain Metastasis publishes 637 high-resolution imaging studies of 75 patients, their respective clinical data, and radiomic features for the cases segmented. Similarly, LUMIERE contains longitudinal glioblastoma MRI with expert RANO evaluation. All data are accessible from the repository Figshare.
Our internal dataset includes over 150 high-resolution multimolecular imaging studies from brain tumor patients, utilizing the ultrafast MRSI technology developed at UIUC, which reveals detailed tissue properties and biochemical profiles. We will also unify data from various centers and protocols to address challenges in intensity, resolution, geometry, and contrast harmonization for subsequent generative model construction.
Aim 2 (Learning-based generative model for disease progression incorporating biological priors and multimodal data): We will develop a novel generative AI framework to model disease progression. This framework will integrate three key components for accurate predictive modeling: (1) multimodal data from qualitative MRI, synthetic quantitative and molecular MRI derived from existing scans, and auxiliary clinical data; (2) biophysical priors of disease progression; and (3) data-driven priors. The integration of these components will utilize advanced conditional generative modeling techniques, specifically the diffusion model.
Aim 3 (Digital twin prototype and evaluation): We will integrate the generative AI framework with uncertainty quantification methods to predict the progression of high-grade glioma. We will assess prediction uncertainties from factors such as image acquisition variations, mapping errors, and inherent disease variability across populations. Additionally, we will develop and validate AI-powered software for predicting glioma growth, with or without treatment, to support optimal, personalized radiation therapy planning.
Timeline: The project timeline spans one year and is divided into four main phases. In Q1 and Q2, we will focus on data collection, translating molecular biomarker data, model theorization, and initial training and evaluation. In Q3, we'll enhance physics-informed training and evaluation while concurrently developing uncertainty characterization methods started in Q2. Finally, Q4 will be dedicated to performance evaluation, assessing the overall effectiveness of the developed systems across all project stages.
Summary of expected outcomes: The proposed project includes innovative components to enhance AI-driven methods for developing digital twins of brain tumor progression. We will validate this by creating a prototype digital twin of brain tumor growth to improve personalized radiation therapy planning. We have identified four specific implementation tasks: Quantitative Biomarkers: We will extract quantitative tissue properties and molecular biomarkers from clinical MRI scans using advanced imaging and AI techniques. Current models rely on MRI data alone, often overlooking early pathological changes. Our synthetic multimodal biomarkers will enhance modeling accuracy by providing detailed insights into tissue properties and molecular signatures. Spatiotemporal Dynamics: We will leverage cutting-edge generative AI, specifically diffusion models, to learn the spatiotemporal dynamics of brain diseases from longitudinal data, significantly improving modeling accuracy. Integrated Framework: We will develop a new AI-powered framework that combines biophysical models with data-driven approaches. This integration will utilize the strengths of both methods to capture complex distributions while addressing challenges like the need for large datasets. Prototype Digital Twin: We will implement a novel digital twin for high-grade gliomas, combining data-driven and biophysical approaches to provide accurate predictions of glioma progression and enhance personalized radiation therapy planning.
Results dissemination: The data collected in this project, along with the developed models and training protocols, will be made available to the research community through open-source platforms (e.g., GitHub), conference presentations, and publications in relevant journals and conferences (e.g., MICCAI).
FAIR and CARE principles: Insights from this study will inform the development of protocols for inclusion/exclusion criteria and standardized data entry procedures to ensure consistent data collection. A centralized database will facilitate data storage and harmonization, with version control for effective monitoring of updates. Comprehensive documentation, including a data dictionary and descriptions of analytical methods, will ensure transparency. We will also create rich metadata for proper indexing in relevant repositories. The de-identified data will be shared on open-access platforms, enabling the broader research community to access it and replicate findings easily.
Understanding and predicting tumor progression has long been a goal for neurologists and oncologists. Recent advancements in neuroimaging and computational sciences have sparked interest in creating digital twins to analyze neurological disorders and personalize patient care. However, two significant technological gaps hinder the development of effective digital twins for tumor diseases: 1) There is currently no technology for routine, high-resolution, noninvasive mapping of brain tissue properties and metabolism. This limits our ability to observe the underlying causes and early pathologies of many neurological and psychiatric conditions, hindering accurate disease modeling. 2) Effective computational tools for accurately modeling the complex spatiotemporal dynamics of brain disease progression are lacking. As a result, existing digital twin technologies have limited clinical impact on early disease prediction and treatment optimization. This project aims to develop an innovative AI-powered computational framework for generative modeling of brain disease progression. The framework will integrate multimodal data, including quantitative tissue properties and molecular biomarkers, to provide comprehensive insights into complex disease processes. By combining biophysical and data-driven priors, it will enhance the accuracy, generalizability, and robustness of disease modeling, serving as a strong foundation for constructing digital twins of tumor disease progression. Creating such a digital twin to model disease progression is vital for modern healthcare. It enables healthcare professionals to predict disease trajectories, optimize treatment planning, and understand the mechanisms of progression. This technology allows for continuous monitoring, providing real-time feedback on treatment effectiveness and facilitating informed decision-making as conditions evolve.
This project is led by an interdisciplinary team at UIUC with expertise in AI, high-performance computing, MRI/MRSI, and brain mapping. Dr. Volodymyr Kindratenko (PI) has applied deep learning to medical image analysis, including a novel transformer-based model, INSTRAS, for segmenting infrared breast images and the YouTubePD benchmark for Parkinson's Disease progression analysis. He has also developed the Diff-Ensembler model for 3D volumetric imaging, addressing inconsistencies in 2D models.
Dr. Zhi-Pei Liang has advanced high-resolution MRSI, leading foundational works in spectroscopic imaging and developing the Spectroscopic Imaging by exploiting spatiospectral CorrElation (SPICE) technology. His lab is now creating frameworks for accelerated and high-resolution MRI/MRSI.
Dr. Yudu Li has focused on AI-driven ultrafast high-resolution MRSI, developing advanced models and algorithms to enhance resolution, SNR, and speed. Recently, he integrated physics-based modeling with deep learning for dynamic X-nuclei MRSI, achieving unprecedented spatiotemporal resolution.
Dr. Shirui Luo has spearheaded numerous research initiatives utilizing deep learning to enhance scientific computation across various fields, including fluid dynamics and remote sensing. Most recently, his focus has shifted to biomedical imaging, where he is exploring innovative applications of deep learning techniques to improve diagnostic accuracy and imaging processes.
To ensure the success of the proposed research project, several key considerations must be addressed: 1) Integration of Multimodal Data: Successfully incorporating diverse data types, such as quantitative tissue properties and molecular biomarkers, is crucial. This integration will provide comprehensive insights into disease progression. 2) Robust Computational Framework: Developing a strong AI-powered computational framework that combines biophysical and data-driven models is essential for enhancing accuracy and generalizability in disease modeling. 3) Validation and Testing: Rigorous validation of the computational models and the digital twin prototype is necessary to ensure their effectiveness and applicability in real-world clinical settings.
