IDDD 2022-2024 Batch

Kushan Raj
B.Tech: Bioengineering
Masters: Complex Systems and Dynamics

BE19B004

A Comprehensive Survey on Diffusion Models and their impact on Medical Imaging
Video

Advisor
: Prof Srinivas Chakravarthy

The integration of advanced image generation techniques in medical imaging holds significant promise for improving healthcare outcomes. This thesis examines the use of Denoising Diffusion Probabilistic Models (DDPMs) in medical imaging to overcome data scarcity and variability. It details the theory behind diffusion models, including innovations like Latent Diffusion Models, ControlNet, and Classifier-Free Guidance. We were able to produce the best-in-class DDPM for medical image generation, beating previous models, though the research is still in the nascent stage. Experimental results demonstrate the capability of these models to generate realistic medical images, potentially transforming medical research and enhancing tumour detection.
Thesis

Placement Adobe | Sarvam AI 


Abhigna Prasad
B.Tech: Bioengineering
Masters: Complex Systems and Dynamics

BE19B011

Reconstruction of ECG signal with an ensemble of hopf oscillators
Video

Advisor: Prof Srinivas Chakravarthy

This study presents a novel approach to model the electrocardiogram (ECG) signal using an oscillatory deep neural network. Understanding the electrical activity of the heart through mathematical modelling is important for diagnosing cardiac disorders. Traditional models become computationally intractable by incorporating too many variables. This work aims to present a simplified model of the heart by reconstructing the ECG signal with a network of coupled Hopf oscillators, allowing parameter identification and mechanistic analysis. The model consists of two stages: an oscillatory stage where oscillator frequencies, connection strengths, and coupling weights are trained via Hebbian plasticity; and a feedforward stage employing a complex-valued multilayer perceptron trained with backpropagation. The power coupling of oscillators enables oscillators with heterogeneous frequencies to be coupled and trained to reconstruct ECG signals. The reconstruction model was evaluated on ECG data from healthy participants to obtain a correlation coefficients up to 0.99 for training data. However, the highly nonlinear nature and large number of parameters limit generalization beyond the training duration. All in all, the model aims to provides a framework for understanding cardiac dynamics and its integration with brain models could offer insights into brain-heart interactions relevant for conditions like depression and epilepsy.

Placement Uniorbit Technologies

Kartik
B.Tech: Bioscience
Masters: Complex Systems and Dynamics

BS19B017

Experimental Data for Visual Search in Parkinson Disease
Video

Advisor: Prof Srinivas Chakravarthy

This thesis explores computational models of decision-making, learning, and action selection, inspired by basal ganglia function, to improve Parkinson's disease research. Using Actor-Critic architectures and Temporal Difference learning, it addresses the complexity of motor control and reinforcement learning, aiming to enhance both fundamental and clinical neuroscience.
Thesis

Placement Axtria

Swarnakshi Kailash
B.Tech: Bioscience
Masters: Complex Systems and Dynamics

BS19B029

Advanced computational framework for epidemic simulation in large scale networks
Video

Advisor Sayan Gupta

This thesis explores an innovative computational framework for simulating epidemics in extensive networks. It addresses the need for accurate models in representing disease transmission dynamics in large, diverse populations. Traditional models like SIR and SEIR lack complexity to capture real-world interactions. Our approach integrates real datasets and advanced computing techniques, particularly CUDA parallelization, to enhance model accuracy and efficiency. We focus on numerical methods, including a novel procedural connectivity method, and explore various epidemic scenarios. Results show the framework's potential to improve epidemic modelling, emphasizing the value of integrating computational power and real-world data in epidemiological modelling.
Thesis

Placement Cleanmax

Deepak Laxman
B.Tech: Bioengineering
Masters: Complex Systems and Dynamics

BE19B019
A gamified tool toi detect Asphasia in patients
Video

Advisor Prof V Srinivas Chakravarthy
Interactive games were developed using Automatic Speech Recognition, Natural Language Processing, and OpenPose to diagnose aphasia. These technologies assess speech, comprehension, and physical task execution, offering a precise, engaging, and patient-friendly alternative to traditional methods. This innovative tool enhances diagnosis and supports improved treatment outcomes for aphasia patients.
Thesis

Sanjay Chandran MM
B.Tech: Bioengineering
Masters: Complex Systems and Dynamics

BE19B030

Chaotic time series prediction and classification using Oscillatory Neural Networks
Video

Advisor Prof V Srinivas Chakravarthy

A trainable Deep Oscillatory Neural Network was developed to predict and classify chaotic time series. Specifically, the model forecasted the third variable at time t based on the values of the other two variables for Rossler and Lorenz systems. Training encompassed chaotic time series forecasting for Rossler, Lorenz, and Mackay-Glass systems. The model demonstrated its efficacy by successfully classifying time series into Chaotic and Periodic categories.
Thesis


Bharat Kumar S
B.Tech: Bioscience
Masters: Complex Systems and Dynamics

BS19B009

Automation of Fugly-Meyer assessment using web-cam feed
Video

Advisor Prof V Srinivas Chakravarthy

The Fugl-Meyer assessment (FMA) are the most widely used tools for assessing upper extremity motor function in stroke survivors. However, due to its subjective and time-consuming nature, it must be directed by therapists in a hospital or clinic environment, and it is not quite suitable for use at home. In this paper, we propose a deep learning-based system for analyzing various FMA tasks. Using a convolutional neural network architecture (Mediapipe), we extract human pose data and use it to generate the necessary features. The FMA scores are obtained using a feature selection approach combined with a Deep learning models based on the FMA’s linguistic guidelines. This framework could be used to assess upper extremity function during stroke recovery to relieve therapists of the time-consuming, repetitive process. It could also be done at home because it does not require any sensors and only uses a computer with a webcam or even a smart phone.

Placement Deloitte - Power Utility consultant(Transmission line - Networks)
Thesis

V Bharathi
B.Tech: Bioscience
Masters: Complex Systems and Dynamics

BS19B031

Hydrodynamics of Cargo-carrying robot for drug-delivery
Video

Advisor Prof Sumesh Thampi

Mathematical modelling of a cargo-carrying robot by re-creating a 2-sphere settling problem and by attaching the two spheres, making one the cargo-carrier and the other the drug/load.

Complex Systems & Dynamics     Indian Institute of Technology Madras     Chennai 600036     India