Research Work

Microscopic medical image analysis is an important area of research in medical imaging. Today's cellular microscopy systems generate volumes of data (BigData). There is a need of using machine learning methods to analyze this data and infer important information out of it, which can help in disease diagnosis. There is also a growing interest in developing robust low cost solutions that are deployable across the board in rural as well as urban areas for clinical diagnosis and treatment. We are working on blood cancers, particularly, white blood cancers, B-ALL and MM. The work is in collaboration with AIIMS, New Delhi. We are building automated tools for disease diagnosis.

Our laboratory focuses on developing advanced artificial intelligence solutions for cardiovascular disease diagnosis and risk prediction. We utilize deep learning architectures to analyze electrocardiogram (ECG) signals for automated cardiac disorder detection. Beyond traditional ECG analysis, our research extends to multimodal approaches, including the use of retinal fundus imaging with models for cardiovascular diagnosis. We are developing the application of graph neural networks for predicting 30-day mortality risk in ST-elevation myocardial infarction (STEMI) patients, incorporating explainability frameworks to provide clinically interpretable insights for healthcare professionals. Our work also encompasses stress level prediction using ECG-derived features, contributing to preventive cardiology and mental health monitoring. Through the integration of interpretability techniques, we ensure our models provide transparent, clinically relevant decision-making support.

Our lab is dedicated to addressing the global challenge of antimicrobial resistance (AMR) through a multidisciplinary approach. Our ongoing work includes the genomic analysis of AMR, to uncover the full range of genetic mutations and host–pathogen interactions that contribute to resistance mechanisms. To combat resistant pathogens, we are designing synergistic drug combinations informed by pharmacogenomic insights, enabling personalized therapeutic strategies. Alongside this, we are advancing computational drug discovery by predicting drug–target affinities at scale to facilitate drug repurposing. Our therapeutic efforts are further strengthened by ongoing investigations into dosage optimization, ensuring both efficacy and safety in treatment development. Cancer genomics investigates the alterations at both the genetic and molecular level that drives tumor development and progression. With the rising cases of breast cancer, a multi-omics approach to understand it is required. We work on this approach to study the impact of chemotherapy in breast cancer by focusing specifically on its long-term impact, such as cardiovascular disease (CVD). By integrating RNA-seq, metagenomic, and metabolomic data, the analysis aims to elucidate how chemotherapy-induced changes in gene expression and gut microbiota contribute to an increased risk of CVD post-chemotherapy. This integrated approach seeks to identify key molecular signatures and pathways that could serve as potential biomarkers or therapeutic targets.

Deep learning models applied in the domain of modern vision and language applications have grown enormously in size, often reaching billions of parameters, making them increasingly difficult to adapt or fine-tune on a single consumer-grade GPU, limiting deployment in data-constrained clinics and on-device settings. We are aiming to develop parameter-efficient compression pipelines that combine low-rank adaptation (LoRA) with mixed-precision training and post-training quantisation to shrink memory requirements, reduce compute use, and preserve performance in vision and language transformer models. At a broader level, we study how to make vision foundation models more portable, affordable, and easy to fine-tune in the medical domain and other resource-constrained environments, enabling wider access to AI-assisted decision support. A movie trailer is a unique selling point for the movie since it serves as the best audio-visual captivation of the scenes, storyline and movie artists. However, it is mostly created by experts making it a traditionally labor-intensive and creative process. Hence there is a strong need for automation due to complexity and cost of manual curation with a wide gap of available commercially viable solutions. Our work aims to generate coherent and visually appealing trailers of any video using a combination of open source tools and in-house designed deep learning / computer vision models.

Diffusion MRI (DMRI) is an emerging non-invasive medical imaging technique primarily used to study white matter structures in brain. Based on the principles of Nuclear Magnetic Resonance (NMR), it utilizes the fact that diffusion of water molecules in white matter region is not isotropic. Presence of nerve fibers, membranes and other macro-molecules creates a diffusion pattern that can be used to study microscopic details and fiber tracts. In DMRI, signals are captured along many diffusion directions that is a time intensive process. Ideally, patients are required to remain stationary during this time while the data is being captured and is very inconvenient. At SBILab, we are working on accelerated reconstruction of DMRI, particularly, HARDI signals using compressive sensing methods via collecting fewer data. The CS-based reconstructed signal is aimed to achieve comparable DMRI representations like ODF and FA with lesser signal samples, thereby decreasing the scanning time. In the near future, we would be working on fibre tractography in HARDI. Some results are also available HERE. Functional Magnetic Resonance Imaging (fMRI) is being used extensively to understand brain function pathways for neuroscience research and clinical applications. It is a widely used imaging modality for representing in-vivo brain state with better spatial resolution compared to electroencephalogram (EEG) and better temporal resolution compared to positron emission tomography (PET). Moreover, it is a non-invasive method. It is a powerful tool to examine various neurosurgical and psychiatric disorders and can also be used, effectively, in the application of pre-surgical planning. In this lab, we are working on the analysis of both fMRI and EEG signals from the point of view of building functional brain networks and resting state networks. In addition, we are working at the problem of accelerated reconstruction of fMRI signals from a fewer measurements using compressive sensing based methods.

Conversational AI is the technology that allows machines to have natural language conversations with users. It utilises methods from natural language processing and machine learning to comprehend, analyse, and answer user inquiries or prompts. These technologies are commonly utilised in virtual assistants, customer service chatbots, and other applications to improve user experience and simplify interactions.

Transform learning (TL) is currently an active research area and is being explored in several applications. It has the advantage that it adapts to signals of interest and is often observed to perform better than the existing sparsifying transforms such as discrete cosine transform (DCT) and discrete wavelet transform (DWT). In general, transform learning is posed as an optimization problem satisfying some constraints, such as transform domain sparsity, that are specific to applications. Joint learning of both the transform basis and the transform domain signal under the constraints renders the optimization problem to be non-convex with no closed form solution. Hence, TL problems are solved using greedy algorithms, wherein a large number of variables are learned. In general, the above requirements and solutions makes TL computationally expensive. Among existing transforms, DWT provides an efficient representation for a variety of multi-dimensional and multiscale signals, capturing signal information into a few significant coefficients. Also, it helps in simultaneous localization of data in both time and frequency. Unlike Fourier transform where the basis is unique, there are a number of wavelets that are either compactly supported or infinitely supported, orthogonal or biorthogonal, discrete or continuous, and so on. In addition, wavelets are connected very naturally with multirate filterbanks that makes them all the more useful in signal processing applications. However, due to the availability of number of wavelets, a question that is always difficult to answer is which wavelet is best suited in a particular application or on a particular signal. This motivates us to explore wavelet transform learning from given signals. We are working on learning DSP hardware friendly wavelets from given signals for applications in inverse problems. Our wavelet transform learning solutions have closed forms leading to fast implementations without the need to look for greedy solution. So far, we have learned dyadic and rational wavelets for both forward as well as inverse problems in signal processing and image processing.

Underwater acoustic channel estimation is a difficult problem under moderate to rough sea conditions. This is primarily due to the fact that the channel is highly non-stationary and is rapidly fluctuating multipath fading channel. In this lab, we are working on channel estimation of underwater acoustic channel employing recent signal processing techniques of sparse sensing, compressed Sensing (CS), etc. In addition, we are working on channel estimation for reliable and robust data transmission in vehicular-to-vehicular (V2V) wireless communication channel. IEEE 802.11p standard is a dedicated wireless V2V standard, which has the biggest challenge of robust channel estimation due to rapid fast fading nature of the channel. Devising efficient signal processing techniques for channel estimation for this scenario is an interesting and challenging problem.

(Yet to be updated.)

To meet the of 5G wireless communication demands of high coverage and seamless user experience, various technologies such as massive multiple input multiple output (MIMO), device to-device communication and mm-wave communication have been proposed. Although these techniques provide an improvement in the performance, they increase the power consumption as well. Therefore, apart from improving the energy efficiency, harvesting the energy from RF signal has gain lot of popularity over renewable resources such as solar, wind, etc. As the RF signal contains both information and energy, therefore, the device can decode the information as well as harvest the energy. As of today, energy harvesting is being explored largely from either solar, wind, or RF signals. We are working on developing newer methods to harvest the energy from analog signals (at any frequency band).