http://dspace.dtu.ac.in:8080/jspui/handle/123456789/85 2026-04-28T03:57:58Z http://dspace.dtu.ac.in:8080/jspui/handle/repository/22672 Title: DESIGN OF EFFICIENT ALGORITHMS FOR BRAIN DISEASE IDENTIFICATION AND CLASSIFICATION Authors: BHATT, KAVITA Abstract: Brain diseases encompass a wide range of disorders, such as neurodegenerative disorders, cerebrovascular disorders, neurodevelopmental disorders, seizure disorders, and brain tumors that impair cognitive, motor, and behavioral functions of human beings. Among these disorders, neurodegenerative disorders are considered the most prominent brain disorders due to their progressive nature, which leads to a continuous decline in cognitive, motor, and behavioral functions. Unlike other brain disorders, these disorders worsen over time and currently have no definitive cure, which makes them a major challenge for healthcare systems worldwide. Alzheimer’s Disease (AD) and Parkinson’s Disease (PD) are the two most prevalent neurodegenerative disorders, affecting millions of people worldwide and imposing substantial social and economic burdens. AD primarily causes progressive memory loss and cognitive decline due to amyloid plaques and tau tangles, while PD mainly leads to motor and non-motor symptoms like tremor, rigidity, and bradykinesia due to dopaminergic neuron loss and Lewy bodies. Both AD and PD are irreversible, progressive neurodegenerative disorders with no particular cure, leading to a gradual decline in cognitive and motor functions and a significant deterioration in patients’ quality of life. Although the progression of these disorders can be slowed with certain medications and therapies. Early and accurate diagnosis is essential to provide timely and effective interventions. Therefore, early detection is crucial for ensuring the much-needed care and support for patients. This thesis aims to develop optimized, automated frameworks for early and ac- curate identification of these brain diseases using biomedical signal analysis and advanced machine learning (ML) and deep learning (DL) techniques. Biosignaling modalities such as electroencephalography (EEG), gait analysis, and speech pattern assessment are cost-efficient and non-invasive in nature. These signals capture essen- tial physiological and behavioral markers reflecting neurological impairments. These iv modalities enable the detection of subtle abnormalities in brain activity, motor func- tion, and communication, providing valuable insights into the onset and progression of neurodegenerative disorders without the need for invasive or expensive clinical procedures. Biomedical signals are typically high-dimensional and contain redundant or ir- relevant information. Identifying the most informative and discriminative features is crucial to improving classification accuracy and computational efficiency. Hence, feature selection-driven optimization models are developed for brain disease detection. For PD, EEG datasets obtained from OpenNeuro are analyzed using statistical and ensemble-based feature selection methods. The Kruskal–Wallis test and Extra tree classifier (ETC) are used to select the most discriminative EEG features. Additionally, a two-stage PD detection framework is developed to enhance diagnostic accuracy and computational efficiency. Initially, an ETC-based feature selection is employed to obtain the most relevant and discriminative speech features while eliminating re- dundant or non-informative ones. These optimal feature subsets effectively reduced dimensionality and improved the model’s ability to capture meaningful variations associated with PD. To address the issue of class imbalance commonly observed in biomedical datasets, the synthetic minority oversampling technique is applied to generate synthetic samples for the minority class. This ensured balanced training data and prevented bias toward the majority class. Then, a stacked ensemble model is em- ployed for classification, which leverages the complementary strengths of individual classifiers. The proposed two-stage framework significantly improved classification performance for PD detection using speech signals. Identifying the most affected brain regions and corresponding EEG channels is crucial for achieving accurate diagnosis and meaningful neuroscientific interpretation in AD. AD causes progressive neurodegeneration that disrupts neuronal connectivity and alters the brain’s rhythmic activity patterns. These abnormalities are not uniformly distributed but are concentrated in specific cortical areas. Therefore, it is essential to analyze EEG signals across multiple lobes: frontal, temporal, parietal, and occipital lobes to identify the dominant brain regions and EEG channels most influenced by v the disease. Identifying these regions enhances the interpretability of ML models, strengthens the physiological validity of classification outcomes, and supports targeted clinical assessments for early and precise AD diagnosis. For this purposes, a Fourier decomposition and Hilbert transform-based EEG signal analysis (FHESA) method is developed. The FHESA method integrates the Fourier Decomposition Method (FDM) and Hilbert Transform (HT) to extract meaningful features from the EEG signal for efficient classification and brain region analysis. The FHESA method aims to efficiently analyze the EEG data to identify the important brain regions vulnerable to AD, and to assess the impact of various EEG channels for the timely and early detection of AD. The accurate detection and classification of neurological disorders is one of the most challenging tasks due to the overlapping clinical symptoms and shared patholog- ical characteristics of diseases such as AD and Frontotemporal dementia (FTD). Both disorders lead to progressive cognitive decline and behavioral impairments, often resulting in misdiagnosis and delayed treatment. Traditional diagnostic methods heav- ily rely on neuroimaging and clinical assessments, which are both time-consuming and costly. Moreover, biomedical signals such as EEG exhibit non-linear and non- stationary behavior, making it difficult for conventional machine learning methods to capture underlying temporal–spectral dependencies. Therefore, there is a need for an algorithm that can extract robust, noise-invariant, and discriminative features capable of representing complex brain activities associated with different neurological conditions. To address this, wavelet scattering transform-based dementia identification and classification (WavDemNet) is proposed. The model leverages the wavelet scatter- ing transform (WST) to extract robust, noise-invariant features that capture essential time-frequency characteristics and a 1-D convolutional neural network (CNN) to learn discriminative patterns for accurate identification and classification of brain diseases. Manual analysis of biomedical signals is time-consuming. To assist clinicians in real-time decision-making, there is a requirement to develop automated and ef- ficient algorithms that can process signals, extract optimal features, and accurately classify neurological disorders with minimal human intervention. For this purpose, an vi automated algorithmic framework is developed for the early diagnosis of PD. The high- resolution superlet transform (SLT) technique is utilized to obtain the time-frequency representation (TFRs) of the signal. SLT employs multiple wavelets to achieve higher TF resolution while being less leaky than a single wavelet, which makes it more sus- tainable to apply to non-stationary signals. In order to identify PD and assess the PD severity rate, the TFRs are fed into deep neural network (DNN) models as input. This approach eliminates the need of additional handcrafted feature extraction methods, as the DNNs are capable of automatically learning hierarchical and discriminative patterns from the TFRs. This model captures signal variations associated with PD progression and results in accurate detection and severity assessment. 2026-02-01T00:00:00Z http://dspace.dtu.ac.in:8080/jspui/handle/repository/22549 Title: DESIGN AND MODELLING OF NANOPHOTONIC DEVICES Authors: DALAL, KIRTI Abstract: This thesis presents a brief overview of the research work carried out on the design and modelling of nanophotonic devices, especially plasmonic switches based on vanadium dioxide (VO2) – a promising phase change material (PCM). Plasmonic switches, which modulate light propagation at the nanoscale, have gained significant attention for applications in integrated photonics, plasmonic logic circuits, and optical communication networks. These devices utilize the tunable plasmonic properties of PCM covered nanostructures under external stimuli — thermal, electrical, or optical — offering ultra-fast switching speeds, wide spectral tunability, compact sizes, and low losses. The thesis also explains the rationale for selecting VO2 and presents an extensive review of VO2 based plasmonic switches developed over the past few decades. Plasmonic switch designs proposed in this thesis are analyzed using Finite Difference Time Domain (FDTD) simulations. Firstly, active broadband plasmonic switches are designed using periodic arrays of metallic nanodisc-dimers with increasing diameters on a gold-coated silicon dioxide (SiO2) substrate, separated by a thin VO2 spacer layer. It is well known that — for a periodic array of circular metallic discs on a substrate, the applied optical field is efficiently coupled into the plasmonic modes of the nanostructure at specific plasmon resonance wavelengths, thus resulting in narrow and strong dips in the reflection spectrum of the nanostructure. Further, there is a red-shift in the plasmon resonance wavelength due to the retardation of the depolarization field when the diameter of the nanodiscs is increased. It can therefore be inferred that by employing an array of dimers of metallic nanodiscs with progressively increasing diameters, multiple wavelengths can be coupled into the plasmonic modes of the nanostructure such that each nanodisc-dimer with a specific value of diameter results in the coupling of incident electromagnetic radiation into the plasmonic modes of the nanostructure at a specific plasmon resonance wavelength. This coupling of incident radiation to plasmonic modes at multiple wavelengths results in an overall broadband resonance dip in the reflection spectrum of the nanostructure, enabling broadband response across the C, L, and U optical communication bands. The proposed designs achieve a broadband extinction ratio (ER) of 5 dB over 650 nm (from 1460 nm to 2110 nm wavelength) and 4 dB over 702 nm (from 1432 nm to 2134 nm wavelength) for a periodic array of five sets of nanodisc-dimers. The trade- off between ER and bandwidth is also explored for design optimization. Next, polarization-independent dual-wavelength switches are proposed using a periodic combination of U and C shaped gold nanostructures on a gold coated SiO2 substrate with a thin VO2 film spacer between the nanostructures and the underlying plasmonic substrate. The U and C shaped nanopillars are placed on the underlying substrate such that there is a spatial offset between them. It is well known that when three nanopillars are arranged in a U shaped nanostructure, two plasmonic modes ⎯ a short wavelength mode and a long wavelength mode ⎯ are generated when the incident light is X-polarized, whereas a single plasmonic mode is generated when the incident light is Y-polarized. For the C shaped nanostructure, this situation becomes inverted, with two plasmonic modes being excited for Y-polarized light and vi one plasmonic mode being excited for X-polarized light. However, when both U and C shaped nanostructures are placed together to form a U-C type plasmonic switch, two plasmonic modes ⎯ a short wavelength mode and a long wavelength mode ⎯ are generated for both X- and Y-polarized incident light. Due to the excitation of two plasmonic modes at both polarization angles of incident light, these U-C type plasmonic nanostructures are employed in this work to realize an efficient polarization-independent multi-wavelength switch by placing them on a VO2 coated plasmonic substrate. The proposed U-C type plasmonic switches exhibit an ER of ~20 dB simultaneously at two wavelengths ⎯ at ~1560 nm and at ~2130 nm ⎯ for a linearly polarized incident light with any polarization angle from 0° to 90°. Further, periodic arrays of identical V-shaped gold nanostructures and variable V-shaped gold nanostructures are designed on top of a gold-coated SiO2 substrate with a thin spacer layer of VO2 to realize multi-wavelength and broadband plasmonic switches, respectively. The periodic array of identical V-shaped nanostructures (IVNSs) with small inter-particle separation leads to coupled interactions of the elementary plasmons of a V-shaped nanostructure (VNS), resulting in a hybridized plasmon response with two longitudinal plasmonic modes in the reflectance spectra of the proposed switches when the incident light is polarized in the x-direction. FDTD modelling is employed to demonstrate that an ER > 12 dB at two wavelengths can be achieved by employing the proposed switches. Further, plasmonic switches based on variable V-shaped nanostructures (VVNSs) — i.e., multiple VNSs with variable arm lengths in one unit cell of a periodic array — are proposed for broadband switching. In the broadband operation mode, an ER > 5 dB over an operational wavelength range > 1400 nm in the near-IR spectral range spanning over all optical communication bands, i.e., the O, E, S, C, L and U bands, is reported. Further, it is also demonstrated that the operational wavelength of these switches can be tuned by adjusting the geometrical parameters of the proposed design. Additionally, the thesis investigates near-field plasmonic switching using pentagon shaped fractal plasmonic nanoantennas placed on a VO2 thin film overlaying a gold-coated SiO2 substrate. These fractal geometries are designed to confine electromagnetic fields to deep sub-wavelength volumes, generating intense field enhancements at the nanogap between the antenna arms. Upon phase transition of the VO2 layer, a significant change in local field distribution is observed, leading to an intensity switching ratio (ISR) exceeding ~2300 for higher fractal orders. The spectral response is shown to be tunable via geometric parameters. These near-field designs hold promise for utilization in areas like surface enhanced Raman spectroscopy (SERS), nonlinear optical phenomena, and nanoscale sensing. In summary, this thesis offers a detailed exploration of VO2 based plasmonic switches, emphasizing their potential for broadband, multi-wavelength, polarization-independent, and near-field switching applications. The proposed designs, supported by rigorous simulations, contribute significantly toward the development of compact and high-performance plasmonic devices for next-generation optical communication and photonic integration. 2025-10-01T00:00:00Z http://dspace.dtu.ac.in:8080/jspui/handle/repository/22546 Title: DEEP LEARNING FRAMEWORKS FOR FACE ANTI-SPOOFING Authors: ANTIL, AASHANIA Abstract: With the rapid integration of facial recognition (FR) systems in access control, banking, and mobile authentication, the risk of face spoofing—also known as presentation attacks (PAs)—has grown significantly. These attacks, carried out using printed images, replayed videos, or 3D masks, threaten the security and reliability of biometric authentication systems. The increasing sophistication of spoofing techniques, combined with the low cost and easy availability of generative tools, underscores the urgent need for robust Face Anti-Spoofing (FAS) or Presentation Attack Detection (PAD) mechanisms. Although several approaches have been proposed, many existing solutions struggle to generalize under challenging conditions involving varied lighting, spoofing materials, backgrounds, and image/video quality. To address these challenges, this thesis proposes a suite of deep learning-based frameworks that are robust, interpretable, and generalizable for real-world face anti- spoofing. The research is structured around four complementary solutions, each targeting a specific dimension of the problem: texture-based learning, multi-modal fusion, spatio-temporal modeling, and generative learning. The first solution introduces a two-stream hybrid framework that fuses handcrafted and deep features to improve spoof detection accuracy. It combines Multi- Level Extended Local Binary Patterns (ELBP) to capture fine-grained texture information with a modified Xception network, enhanced by Squeeze-and-Excitation (SE) blocks for channel-wise feature reweighting without increasing complexity. This design balances expressive power and computational efficiency, enabling the model to handle diverse spoofing conditions and maintain generalization across datasets. The second solution, MF2ShrT, addresses multi-modal fusion by leveraging the power of Vision Transformers (ViTs). It uses overlapping patches to emphasize local contextual cues and introduces SharLViT, a shared-layer transformer backbone that improves feature representation while reducing parameter count. A novel T-Encoder- viii based Hybrid Feature Block is employed to mine inter-modal dependencies across RGB, depth, and IR streams. The Adaptive Weighted Fusion and Classification Block (AWFCB) then learns to dynamically combine these features, emphasizing salient cues while suppressing redundant information—resulting in a flexible and accurate spoof detection system. The third solution focuses on the temporal dimension of FAS by proposing Bi- STAM, a Bi-Directional Spatio-Temporal Adaptive Modeling framework. Aiming to capture motion inconsistencies and subtle dynamics in video-based attacks, it introduces two key components: a Temporal Adaptive Block (TAB) to balance motion and static information, and a Spatial Adaptive Block (SAB) to enhance texture representation while filtering noise. These are fused via a Feature Aggregation Block (FAB) to yield a unified spatio-temporal representation, significantly boosting generalization and performance on video-based spoof detection tasks. The fourth solution, PolarSentinelGAN, presents a novel generative adversarial framework that enhances spoof classification through depth map generation. By fusing RGB and Multi-Scale Retinex with Color Preservation (MSRCP) inputs, the model uses Dual Polarized Attention (DPAttn) to focus on discriminative regions. A dedicated Feed Forward Block (FFB) within the generator facilitates the transmission of rich features, while optimized latent variables improve generalization across attack types and datasets. All four frameworks are extensively evaluated using standard intra- and cross- dataset testing protocols on public benchmarks, and are further supported with explainability techniques such as class activation mapping and feature occlusion testing. The results demonstrate strong real-time performance, robustness, and scalability. The proposed methodologies in this thesis makes substantial contributions toward the development of next-generation face anti-spoofing systems. The proposed methods not only address key challenges in generalization, efficiency, and interpretability, but also pave the way for practical deployment in critical domains such as finance, border security, surveillance, and consumer electronics. Future directions include exploring ix federated learning, privacy-aware architectures, and continual domain adaptation to further enhance system reliability in dynamic real-world environments. 2025-12-01T00:00:00Z http://dspace.dtu.ac.in:8080/jspui/handle/repository/22545 Title: DESIGN AND SIMULATION OF NOVEL NANO-SCALE DEVICES FOR BIOSENSOR APPLICATIONS Authors: KUMAR, ANIL Abstract: Field Effect Transistor (FET)-based biosensors have become a cornerstone in biomedical diagnostics, environmental surveillance, and biosecurity because of their exceptional sensitivity, fast response time, and compact design. These biosensors utilize the intrinsic characteristics of FETs to identify biological and chemical sub- stances by observing variations in electrical parameters such as current or voltage. Among the various FET-based biosensors, silicon nanowire (SiNW)-based biosen- sors are particularly noteworthy. The substantial surface-to-volume ratio of SiNWs greatly amplifies their interaction with target biomolecules, enhancing sensitivity and lower detection limits. Furthermore, the compatibility of SiNWs with existing semi- conductor fabrication technologies facilitates their integration into complex sensing systems. This thesis investigates various FET-based devices, including Charge Plasma Tun- nel FETs, Non-Reconfigurable Silicon Nanowire (SiNW) FETs, and Reconfigurable SiNW FETs, focusing on their promising uses in biosensing. Initially, a sensitivity analysis of CP Tunnel FETs (CP-TFETs) was conducted by optimizing cavity place- ment and size to boost biosensor performance. The study then focuses on biosensors utilizing Silicon Nanowire FETs, examining reconfigurable and non-reconfigurable types. This evaluation covers calibration, sensitivity parameter analysis, and a per- formance comparison with cutting-edge biosensors. Additionally, Spacer Engineering techniques are applied to enhance the performance of RFET-based biosensors. Fol- lowing this, noise and sensitivity analysis was performed, assessing distortion and linearity. An analytical model was formulated for RFET-based biosensors and its sensitivity compared to leading biosensors. Below, comprehensive assessments of the viii effectiveness of each device for biosensing are provided. Optimizing Cavity Position in the Charge Plasma Tunnel FET-Based Biosensor: We performed a comparative analysis of different cavity positions (source, gate, and drain) in CP-TFET for biosensor applications. The intrinsic properties of biomolecules, such as the dielectric constant and charge density, are leveraged to detect biomolecules within the nanogap cavities. Our findings indicate that the placement of cavities significantly impacts the device’s sensitivity and electric field distribution, suggesting optimal configurations for enhanced biosensing performance. Performance Evaluation of Reconfigurable FET (RFET) and Non-Reco- nfigurable FET for Biosensor Application: We assessed the SiNW FET-based biosensor with and without reconfigurable features and examined the sensitivity of the proposed biosensors. Furthermore, we compare the sensitivity of our proposed device with that of the advanced FET-based biosensors. High-Performance RFET-based Biosensor using Spacer Engineering: We explored the Spacer Engineering Reconfigurable Silicon Nanowire Schottky Bar- rier Transistor (SE R-SiNW SBT) as a label-free biosensor capable of detecting dual- polarity biomolecules with high sensitivity, selectivity, and linearity. The device’s dual-gate configuration significantly enhances the modulation of the Schottky tun- neling width and channel potential, providing a robust framework for biosensing ap- plications. Noise and Sensitivity analysis of the RFET for Biosensor Application: This work presents a comprehensive noise and sensitivity analysis of the proposed RFET tailored for biosensor applications. Experimental calibration corroborates the simulation results, showcasing the device’s exceptional sensitivity and noise char- acteristics. The findings underscore the proposed RFET’s potential for real-time, low-power biosensing applications. ix Analytical Modeling of the RFET for Biosensor Application: The analyt- ical modeling of the proposed RFET biosensor, featuring a cavity under the control gate, is introduced to further advance biosensing capabilities. This design leverages the dielectric modulation effect to achieve high sensitivity and selectivity in detect- ing biomolecules. The device’s architecture and operation are optimized to enhance biosensing performance, showcasing significant potential for various biosensing appli- cations. 2025-12-01T00:00:00Z