RESEARCH

MRI Super-Resolution with Deep Learning (FSRCNN & SRDenseNet)

A major challenge in clinical diffusion MRI is the trade-off between image resolution, acquisition time, and scanner hardware cost. High-field MRI scanners (7T) are largely unavailable in resource-limited settings, yet their superior image resolution is critical for detecting subtle microstructural changes in TBI, glioma, and prostate cancer.

We leverage Fast Super-Resolution Convolutional Neural Networks (FSRCNN) and SRDenseNet architectures to enhance the spatial resolution of diffusion MRI — without requiring expensive high-field hardware. These AI models learn to reconstruct high-resolution MRI images from low-resolution inputs, achieving near-3T image quality from standard 1.5T acquisitions.

Key contributions of this work:

Applying FSRCNN and SRDenseNet to diffusion MRI from TBI, glioma, and prostate cancer datasets
Evaluating super-resolution performance using SSIM, PSNR, and downstream diagnostic metrics
Combining super-resolution with advanced diffusion models (NODDI, DKI) for enhanced tissue microstructure characterization
Developing a clinically deployable pipeline suitable for standard 1.5T and 3T MRI scanners

This approach establishes a novel benchmark in TBI diagnostics and broadens access to advanced diffusion MRI analysis in low-resource clinical settings across India and the broader Asian region.

Hybrid Multi-dimensional MRI (HM-MRI) for Tissue Composition Mapping in Biopsy-Naive Prostate Cancer

Conventional multiparametric MRI acquires T1-weighted, T2-weighted, and diffusion-weighted contrasts in separate sequences — increasing scan time and making direct multi-parametric tissue characterization difficult. Hybrid Multi-dimensional MRI (HM-MRI) overcomes this by simultaneously encoding T1, T2, and diffusion contrasts within a single acquisition framework, enabling comprehensive, co-registered tissue composition maps of the prostate gland.

Our ongoing work is the first study of its kind in the Asian population. We aim to:

Generate quantitative tissue composition maps (water fraction, lipid fraction, fibrosis fraction) of the prostate non-invasively
Validate HM-MRI-derived tissue composition against whole-mount pathology from prostatectomy specimens
Use HM-MRI to guide targeted prostate biopsy — addressing multi-focality and molecular heterogeneity of PCa
Compare HM-MRI with conventional mpMRI (PI-RADS) for detection of clinically significant prostate cancer
Develop AI-assisted analysis of HM-MRI data for automated lesion detection and risk stratification

This study addresses critical clinical challenges of PSA-based overdiagnosis and errors associated with software-based fusion biopsy devices, potentially transforming how prostate cancer is detected and characterized in clinical practice across India.

This will be the first study of its kind in the Asian population.

qMRI and AI Modeling for Early Diagnostics and Prognostic Monitoring in Brain Tumors

Brain tumors, particularly gliomas, represent a significant clinical challenge due to their aggressive nature, heterogeneous biology, and poor prognosis. Early detection and accurate prognostic monitoring are critical for improving patient outcomes, yet current imaging techniques often fall short in providing quantitative, real-time insights into tumor metabolism, microenvironment, and response to therapy. Building on advancements in quantitative Magnetic Resonance Imaging (qMRI) and artificial intelligence (AI), this project aims to develop novel MRI-based biomarkers and integrated AI models for glioma management.

We are utilizing ML/DL algorithms (e.g., deep neural networks and convolutional models) to analyze multimodal quantitative MRI (qMRI) data. AI will predict tumor progression, differentiate low- from high-grade gliomas, and predict survival outcomes by processing features like apparent diffusion coefficient (ADC), cerebral blood volume (CBV), and contrast enhancement patterns.

AI-Based Single-Shell to Multi-Shell Diffusion MRI Synthesis

Advanced diffusion MRI models such as NODDI (Neurite Orientation Dispersion and Density Imaging), CHARMED, and MAPMRI require multi-shell acquisitions — data collected at multiple b-values (e.g., b=0, 1000, 2000, 3000 s/mm²) with many diffusion directions. While these models provide far richer tissue microstructure information than standard DTI, multi-shell acquisitions are time-consuming and not routinely available in standard clinical protocols.

We develop AI-based synthesis models — including Generative Adversarial Networks (GANs) and score-based diffusion generative models — to synthesize full multi-shell diffusion MRI datasets from standard single-shell clinical acquisitions (typically b=1000 s/mm² with 32 directions).

This work enables:

Computation of NODDI-derived metrics (ICVF, ODI, ISOVF) from single-shell data — without protocol changes
CHARMED and MAPMRI microstructure modeling on standard clinical acquisitions
Improved white matter characterization in TBI, glioma, and Alzheimer's disease
Retrospective analysis of large clinical dMRI databases acquired without multi-shell protocols
Significant reduction in scan time and cost while maintaining advanced modeling capability

This methodology bridges the gap between the richness of advanced diffusion modeling and the constraints of routine clinical MRI — making state-of-the-art microstructure imaging accessible at scale.

Integrating Radiomics, Metabolomics, and Radiogenomics with Deep Learning for Renal Tumor Analysis

Introducing a groundbreaking project that combines the power of radiomics, metabolomics, and radiogenomics signatures with state-of-the-art deep learning techniques to revolutionize the field of renal tumor analysis. Our innovative approach aims to extract comprehensive data from medical imaging, metabolite profiles, and genomic information to uncover hidden patterns and insights in renal tumors.

By integrating deep learning algorithms, we can unlock the full potential of these multi-dimensional datasets, enabling more accurate diagnosis, personalized treatment strategies, and improved patient outcomes. The synergy of radiomics, metabolomics, and radiogenomics holds tremendous promise for enhancing our understanding of renal tumors and guiding clinical decision-making.

With this project, we aim to push the boundaries of medical imaging and computational analysis, paving the way for a new era in renal tumor research and patient care.

Assessing Morphometric, Metabolic, and Tractographic Alterations in Alzheimer's Disease using mpMRI with Integrated Transfer Learning

Our lab has also started exploring the potential of MR methods in various Alzheimer's disease (AD). AD is a progressive neurodegenerative disorder characterized by cognitive decline and memory loss. This study aims to investigate the specific cortical and subcortical changes, metabolic changes, and tractography changes in AD using Multiparametric MRI (mpMRI) techniques and Transfer Learning techniques. mpMRI can provide quantitative measurements of the brain's microstructure, which can reveal subtle changes in the brain's organization and function as transfer learning focuses on leveraging knowledge gained from one task to improve performance on a related task.

mpMRI scans will be performed using a 1.5T/3T MRI scanner. Cognitive function will be assessed using standardized neuropsychological tests. In house modified, Freesurfer, LC Model, DSI studio and other software will be used for comprehensive analysis of mpmRI of AD. The study is expected to provide insights into the specific alterations that occur in cortical and subcortical structures such as reduced hippocampal, entorhinal cortex volumes, and cortical thickness. Further, we expect changes in FA values in different regions of brain particularly in hippocampus in AD. We also expect metabolic alterations NAA/myo-inositol in AD. These alterations can be used as biomarkers for AD diagnosis and monitoring disease progression.

Furthermore, incorporating transfer learning enhances the accuracy and predictive power of the analysis, and can offer a cutting-edge approach to understanding the intricacies of the disease progression. This combined approach can offer a holistic understanding of AD, paving the way for enhanced diagnostic procedures and potential therapeutic interventions.

Traumatic Brain Injuries: Advanced Diffusion MRI Techniques and AI Methods

India faces significant challenges in managing Traumatic Brain Injury (TBI), with current diagnostic tools like CT scans and Diffusion Tensor Imaging (DTI) demonstrating limitations in specificity and accurate assessment of TBI including assessment of crossing fiber tracts in TBI, highlighting a need for more advanced mathematical and diffusion approaches

To address this, our pioneering approach integrates advanced multicompartment diffusion-MRI models (NODDI, DKI, etc) and groundbreaking AI/ML methods (FSRCNN, SRDenseNet) to enhance the resolution of MRI scans. This circumvents the need for costly high-field machinery, establishing a novel benchmark in TBI diagnostics.

Advancing Breast Cancer Diagnosis and Treatment through AI and mpMRI Techniques

Breast cancer is the most frequently diagnosed cancer, accounting for 11.7% of all new cancers diagnosed in 2020. Breast cancer is a major healthcare challenge around the world, necessitating early detection and effective treatment strategies to improve patient outcomes. Recent advancements in artificial intelligence (AI) based methods (eg. radiomics, deep, and machine learning) have shown potential to improve the diagnostic accuracy in improved detection and characterization of breast lesions.

We evaluate the potential of mpMRI methods for suspicious clinical symptoms and equivocal findings on mammography and detection of additional tumor foci in the ipsilateral and contralateral breast.

Design and Development of Computer Vision-based Screening Tool for Automatic Diagnosis of Osteoporosis (Co-Investigator)

Osteoporosis is a chronic disease characterized by progressive bone loss and changes in microstructural levels, resulting in increasing bone weakness and fragility. The main objective of this research project is to develop a “Computer Vision based Screening Tool for Automatic Diagnosis of Osteoporosis” to assess the bone health condition and its severity using artificial intelligence (AI) based methods. This device will be economical, user friendly and it will not require any expert or specialized manpower to operate it.

The resulting prototype shall be practically usable and affordable, making it well suited for rapid medical screening at primary health care centers.

Other areas:

Beside these, we have also initiated a preliminary work on emerging MR methods (MRS, rfMRI, etc) in neurological diseases and NAFLD.

Page updated

Google Sites

Report abuse