Ekansh Chauhan

I am Ekansh Chauhan, currently pursuing a Master's by Research at the International Institute of Information Technology, Hyderabad, affiliated with the Cancer Diagnostics (DIAG), Centre for Visual Information Technology (CVIT) lab. Under the guidance of professors Vinod P.K. and C.V. Jawahar, my research focuses on the automated classification of tumors using histopathology images. In this project, my work involves collecting data from affiliated institutions and employ deep learning methods to detect cancer tumors.

Before I joined IIIT Hyderabad, I was a Research Fellow at iHub Data, where I worked with Prof. Bapi Raju S on developing a low-parameter model, LRH-Net, designed to detect multiple fatal cardiovascular diseases at once in resource-constrained environments. Also, through proposed multi-level knowledge distillation, I was able to reduce the required input leads to two or three (generally twelve), enhancing its user-friendliness and suitability for edge devices.


I completed my Bachelors from GGSIPU University Delhi in Information Technology. I worked in the areas of evolutionary algorithms, detection and prediction of COVID-19 using deep learning, and interned at The IIT-BHU - Indian Institute of Technology (BHU) Varanasi with Prof. Hari Prabhat Gupta on quality assessment of River Ganga using machine learning.

Research Interests: My research interests lie broadly at the intersection of computer vision and healthcare. My goal is developing autonomous healthcare diagnosis systems which integrate diverse modalities and enhance patient outcomes, including predicting disease progression and identifying treatments, with a focus on explainability and edge device applicability.


2022 - Current
2023 - 2024
2021 - 2022
2019 - 2023
2020
2017-2021

Selected Publications


Multiple Instance Learning for Glioma Subtype, Grading and IHC biomarkers using H&E Stained WSIs: An Indian Cohort Study

Ekansh Chauhan*

,

Amit Sharma, Megha S Uppin, C.V. Jawahar & P.K. Vinod

MIDL'24

Paper / Code

This study introduces advancements in brain tumor (glioma) management through multiple-instance learning, ResNet-50 for feature extraction, and the DTFD feature aggregator for histopathology analysis. It sets new performance benchmarks for glioma subtype classification, achieving AUCs of 88.08 on the Indian demographic dataset (IPD-Brain: in-house) and 95.81 on the TCGA-Brain dataset. Additionally, it establishes benchmarks for grading and identifying immunohistochemical molecular biomarkers in gliomas using H&E stained slides. The research demonstrates a significant alignment between the model's decisions and pathologists' diagnostic methods, showcasing its potential to replicate professional diagnostic procedures.

* Gif Source
Lupus Nephritis Subtype Classification with only Slide Level Labels

Amit Sharma

,

Ekansh Chauhan*

,

Megha S Uppin, Liza Rajasekhar, C.V. Jawahar & P.K. Vinod

Submitted

Paper / Code

Our research introduces LupusNet, a novel deep learning model designed for lupus nephritis classification using only slide-level labels, avoiding the need for detailed glomerular-level labeling. We created the largest multi-stained digital histopathology dataset for lupus nephritis from the Indian population. LupusNet leverages a multiple instance learning approach and has demonstrated high effectiveness, with an AUC score of 91.0%, F1-score of 77.3%, and accuracy of 81.1% in identifying membranous and diffused lupus nephritis subtypes.

LRH-Net: A Multi-level Knowledge Distillation Approach for Low-Resource Heart Network

Ekansh Chauhan*

,

Swathi Guptha, Likith Reddy & Bapi Raju

MICCAI Workshop

Paper / Code / Video (Coming Soon)

We introduce LRH-Net, a low-parameter model for ECG anomaly detection in resource-constrained settings. Leveraging multi-level knowledge distillation (MLKD), LRH-Net benefits from insights distilled from higher-parameter teacher models trained on diverse lead configurations. With 106 fewer parameters and 76% faster inference than the teacher model, LRH-Net achieves efficient detection of cardiovascular diseases, exhibiting a scaled performance of 3.25% on reduced lead data, making it well-suited for deployment on edge devices.


News & Announcements