Assoc. Prof. Bala Murugan MS
Vellore Institute of Technology, India
Dr. Bala Murugan is an Associate Professor at the School of Electronics Engineering, VIT University, Chennai, with over 19 years of experience bridging academia and industry. He holds a prestigious international leadership role as the Chair of the ITU-T Focus Group on AI for Agriculture (FG4IA) in Geneva. Dr. Murugan maintains strong research ties with Japan, having served as a two-time Visiting Researcher at Kyoto University (WPI-iCeMS) and an Invited Instructor at Waseda University. His research expertise lies at the intersection of Artificial Intelligence, Computational Biology, and Biomedical Engineering. He specializes in developing advanced machine learning models for cellular senescence research and medical diagnostics. Dr. Bala Murugan has spearheaded multiple industrial consultancy projects in healthcare technology and IoT. A globally recognized speaker, he has delivered keynotes across India, Singapore, Japan, and Silicon Valley, advocating for the integration of intelligent algorithms into bio-digital systems to revolutionize modern healthcare.
Speech Title: "Harnessing State Space Models (Mamba Architecture) for Early-Stage Lung Cancer Detection"
Abstract: Lung cancer remains a leading cause of cancer-related mortality, making the early detection of pulmonary nodules in Computed Tomography (CT) scans critical. Traditional Deep Learning approaches, such as 3D Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), often struggle with the computational bottleneck of processing high-resolution 3D volumetric data. Transformers, in particular, face quadratic complexity with respect to sequence length, limiting their efficiency in modeling global contexts across numerous CT slices. This talk introduces the application of Mamba, a novel architecture based on State Space Models (SSMs), which offers linear scaling with sequence length. Utilizing the LIDC-IDRI public dataset, we demonstrate how Mamba can effectively treat 3D CT volumes as long sequences of 2D slices. This approach allows for the efficient capture of long-range dependencies and subtle nodule patterns without the high computational overhead of traditional attention mechanisms. The session will explore how this architecture improves both inference speed and diagnostic accuracy for early-stage screening. Also if there is any flyer of the conference please do share so it helps disseminating across my network.
