Sparse Diffusion MRI reconstruction and Tractography
This research focuses on reconstructing high-quality brain connectivity from sparsely sampled diffusion-weighted imaging (DWI) data using deep learning techniques. Traditional dMRI acquisition requires dense sampling across multiple diffusion directions, making it time-consuming and less feasible in clinical settings. Our work significantly reduces scan time while preserving accuracy, enhancing the efficiency and scalability of brain imaging pipelines.
We developed SwinDTI, a transformer-based neural network that estimates diffusion tensors and enables reliable fiber tractography from limited DWI inputs. Unlike conventional model-based methods (e.g., Unscented Kalman Filters), our pipeline uses data-driven neural models to infer fiber directions, allowing for end-to-end learning and better generalization across subjects and scanning conditions.
Key highlights:
Enables fast and accurate tractography from sparse DWI, reducing burden on both scanners and patients.
Supports early diagnosis of neurodegenerative diseases such as Alzheimer’s and Frontotemporal Dementia.
Integrates deep neural networks to replace hand-crafted models for fiber orientation estimation.
Developed in collaboration with Shiv Nadar IoE and Harvard Medical School researchers.
Collaborator:
Dr. Rajeev Kumar (Shiv Nadar IoE)
Students:
Abhishek Tiwari(PhD, Graduated in 2024), Ananya Singhal(BTech, Graduated in 2024)
PhD Thesis of Abhishek: Deep Learning Based Framework for DTI Parameters
Estimation and Analysis for Sparse Diffusion MRI data
Unsupervised Image Segmentation using Graph Neural Network
This research explores a novel approach to image segmentation by leveraging the structural strengths of Graph Neural Networks (GNNs) in an unsupervised setting, specifically targeting both medical imaging and natural image datasets.
We proposed UnSegMedGAT and UnSeGArmaNet, two state-of-the-art GNN-based models that operate without pixel-level labels. Our architecture constructs a graph over image superpixels, where nodes represent regions and edges encode spatial and appearance similarities. The segmentation is driven by a modularity-based loss function that encourages the discovery of coherent and semantically meaningful regions within the image.
Key highlights:
No need for labeled data, enabling applications in domains where annotations are scarce or expensive (e.g., medical imaging).
Incorporates attention mechanisms and graph pooling to capture global and local context.
Outperforms baseline clustering and GCN-based models on standard segmentation benchmarks.
Demonstrated strong generalization across datasets (e.g., medical scans, PASCAL VOC, COCO sub-samples).
Collaborators:
Dr. Nitin Kumar, Dr. Snehasis Mukharjee(Shiv Nadar University)
Students:
Kovvuri Sai Gopal Reddy(BTech, Graduated in 2024), Bodduluri Saran(BTech, Graduated in 2024), Mudit Adityaja(BTech, Graduated in 2024)
Shape Analysis on Anatomical Changes of Brain Subcortical Structures in Neurodegenerative Diseases
During my Ph.D. research, I focused on developing advanced shape analysis methods to study morphological
changes in subcortical brain structures associated with neurodegenerative diseases such as Alzheimer’s and Parkinson’s.
I recognized that traditional volumetric analyses often miss subtle yet clinically meaningful deformations, so I concentrated on
capturing fine-grained geometric variations in regions like the hippocampus, amygdala, caudate, and thalamus.
One of my key contributions was designing a hierarchical generative modeling framework in Riemannian shape space, utilizing a Monte Carlo
Expectation-Maximization (EM) algorithm. This framework allowed for statistically robust hypothesis testing on shape variations,
enabling more precise detection of disease-specific morphological patterns. I presented this work at the Medical Image Computing and
Computer Assisted Intervention (MICCAI) conference in 2016.
Later in my research, I explored the use of deep neural networks guided by shape priors to improve object segmentation performance—especially
when working with limited or noisy training datasets. This approach proved effective in segmenting subcortical brain structures from MRI scans
and contributed to more accurate assessments of disease progression. I presented this work at the IEEE International Symposium on
Biomedical Imaging (ISBI) in 2020.
Through this line of work, I aimed to develop more sensitive and reliable imaging biomarkers for the early detection and
monitoring of neurodegenerative disorders. By combining statistical modeling with deep learning techniques, my research contributes
to improving the clinical utility and interpretability of neuroimaging pipelines.
My Thesis: Hierarchical Pointset-Based Statistical Shape Modeling
and Applications
PhD Defense Presentation
