DSpace Community:

STRUCTURED LATENT SPACE EXPLORATION WITH TRANSFORMER ENCODERS FOR DIVERSIFIED AND PERSONALIZED MULTIRATER MEDICAL IMAGE SEGMENTATION

2026-05-01T00:00:00Z

Title: STRUCTURED LATENT SPACE EXPLORATION WITH TRANSFORMER ENCODERS FOR DIVERSIFIED AND PERSONALIZED MULTIRATER MEDICAL IMAGE SEGMENTATION Authors: SHUKLA, KESHU; Verma, Bindu (SUPERVISOR) Abstract: Multi-rater medical image segmentation requires models that capture inter-annotator disagreement, not average it away. Standard probabilistic models process all rater annotation through one shared encoder: when four radiologists label the same nodule differently, their reconstruction gradients partially cancel inside that encoder and the latent code ends up as a gradient-weighted compromise across all four boundary deci sions. This is why prior samples from such a model cluster near the mean annotation rather than spanning the actual range of what qualified radiologists drew. We address this by replacing the shared posterior with N independent per rater posterior encoders qi(zi | x,yi), one per annotator. Each receives a 2-channel input: the image and a single rater mask. Gradient isolation follows from the per-rater ELBO decomposition, not from any regularisation: by the chain rule, ∂Li/∂zj = 0 for i ̸ = j, so rater i’s reconstruction gradient cannot reach rater j’s encoder. On LIDC-IDRI (1,018 CT scans, 4 radiologists, 1,609 nodule patches, 4-fold cross-validation), the per rater model (Stage 1 only) achieves GED 0.1444±0.0141 (−4.2%) and Dice_match 0.9112±0.0061 (+2.28% relative) over the full D-Persona two-stage pipeline. A systematic ablation tests transformer-based encoders (MiT-B2), orthogonality regu larisation, a discretised prior bank (k = 100), a dual diversity loss, and Stage 2 style vectors against the D-Persona baseline. Per-rater posteriors are the only modification that consistently improves both metrics at once. Transformer encoder capacity, tested as a direct competing architectural hypothesis, does not resolve the training-level gradient conflict. Dice_soft is unchanged at 0.9015: the gain comes from improved diversity and per-rater accuracy, not from higher average prediction quality. We test the model’s behaviour when not all annotators label every training image (the common clinical situation in multi-rater datasets). Under full sparsity (one annotator), the shared baseline undergoes gradient collapse: mean pairwise cosine similarity of reconstruction gradients rises from 0.167 (full annotation) to 0.976; within fold standard deviation shrinks approximately 19-fold (0.439 → 0.023). Per-rater posteriors maintain zero alignment by construction in all sparsity levels. The GED advantage grows with sparsity: +11.5% with three annotators, +17.8% with two, iv +21.4% with one. All 12 per-fold comparisons favour the per-rater model (sign-test p <0.001). At full annotation, both models are statistically equivalent (0.5% gap, within noise); the advantage is tied to sparsity, not general accuracy. On NPC-170 (170 nasopharyngeal carcinoma MRI cases, 4 annotators), the GED difference is 0.0011, within seed variance ±0.0085. The method works on a different anatomy and dataset. A third contribution analyses inter-rater annotation disagreement on 1,603 LIDC-IDRI cases using the nine per-rater clinical attribute ratings. Nodule margin clarity is the strongest predictor of inter-rater mask variance (Pearson r = 0.318, p < 0.001, confirmed across all four folds independently), followed by lobulation (r =0.243) and texture (r = 0.210). Malignancy is negatively correlated with mask variance (r =−0.202, p<0.001); a nodule rated highly suspicious need not have an ambiguous boundary, and one with an unclear margin need not look malignant. These findings point to where uncertainty-aware segmentation matters most: ill-defined, lobulated, part-solid nodules.

EFFICIENCY-DRIVEN SINGLE IMAGE SUPER-RESOLUTION USING ATTENTION-ENHANCED RESIDUAL FEATURE DISTILLATION NETWORKS

2026-05-01T00:00:00Z

Title: EFFICIENCY-DRIVEN SINGLE IMAGE SUPER-RESOLUTION USING ATTENTION-ENHANCED RESIDUAL FEATURE DISTILLATION NETWORKS Authors: MANGAL, ISHAN; Verma, Bindu (SUPERVISOR) Abstract: Image reconstruction based on low-resolution input appears to be a straightfor ward task but it is fundamentally ill-posed since there may exist many possible solu tions to the problem– all high-resolution images that could correspond to the observed downsampled version. The existing approaches in DL generate outstanding outputs; however, almost all of them require large amounts of computational power, which is incompatible with real-time inference on mobile devices, edge computing units, and camera hardware. ARFD-ESPCN is an image super-resolution architecture that builds upon the ES PCN sub-pixel upsampling structure but utilizes Feature Distillation Blocks, Squeeze and-Excitation channel-wise attention, and Global Feature Fusion layer. The key fea ture of the proposed architecture lies in preserving the speed of ESPCN by using only low-resolution convolutions and applying PixelShuffle once, at the final layer. This design replaces ESPCN’s shallow encoder-decoder with deeper and more competitive architecture that exploits six FDBs for splitting and merging convolution features with higher attention on the aspects not caught previously. The final model has 597,904parameters, needs 4.87 GFLOPsfora640×360input, and runs in 4.78 ms on a mid-rangeGPU.TrainingusedtheDF2Kdataset(DIV2Kplus Flickr2K, roughly 3,450 images) with L1 loss for 720 epochs and Charbonnier loss for the remaining 80, combined with flip and rotation augmentation. The optimiser is Adam with a cosine-annealing schedule that warms up over the first 50 epochs and decays toward 10−6 by the end. On standard benchmarks the model scores 31.57 dB / 0.8861 SSIM on Set5, 27.21 dB / 0.7447 on Set14, 26.22 dB / 0.7029 on BSD100, and 25.37 dB / 0.7606 on Urban100. That beats the original ESPCN by over 2 dB on Set5 and matches VDSR while using roughly 125× fewer floating-point operations. Compared with heavier efficient models like IMDN and RFDN the quality gap is about 0.6–0.7 dB, but ARFD-ESPCN runs 2–3× faster. A step-by-step ablation decomposes the total 2.11 dB gain into individual contribu tions: the training schedule accounts for 1.04 dB (the largest single factor), DF2K data adds 0.45 dB, SE attention with residual connections contributes 0.38 dB, augmenta tion plus L1 loss gives 0.15 dB, and the distillation block structure adds 0.07 dB. The practical lesson is that for models under one million parameters, getting the training pipeline right matters at least as much as architectural design. The model is small enough to fit on mobile accelerators and fast enough for 60 fps video pipelines, making it relevant for medical imaging, satellite photo enhancement, and streaming scenarios where latency budgets are tight. Possible extensions include real-world degradation handling via BSRGAN-style pipelines, perceptual and adver sarial losses for sharper visual textures, window-based spatial attention for periodic patterns, and INT8 quantisation for hardware without floating-point units.

DESIGN AND DEVELOPMEN T OF SECURE FEDERATED LEARNING FRAMEWORK FOR DATA-SENSITIVE APPLICATIONS

2026-02-01T00:00:00Z

Title: DESIGN AND DEVELOPMEN T OF SECURE FEDERATED LEARNING FRAMEWORK FOR DATA-SENSITIVE APPLICATIONS Authors: NARULA, MANU; Vishwakarma, Dinesh Kumar (SUPERVISOR); Meena, Jasraj (CO-SUPERVISOR) Abstract: In an era where security breaches and cyber fraud have become increasingly common, con- cerns about data privacy have also risen exponentially. This presents a substantial hur- dle for training Artificial Intelligence (AI) applications that handle sensitive data, such as healthcare and finance. Federated Learning (FL) provides a solution for Data-Sensitive Applications (DSA), enabling collaboration between competing parties while guaranteeing complete control over data privacy. Over time, the attacks on FL have evolved, and the native technology alone is insufficient to maintain the expected privacy levels when han- dling such data. Hence, efforts are made to integrate other technologies with FL to enhance certain aspects, such as resource management, security, privacy, and efficiency, with varying results and pros and cons. This thesis examines the integration of security mechanisms with FL, focusing on networks with resource scarcity, which include AdHoc networks, Internet of Things, and Internet of Medical Things (IoMT), commonly found in the field of healthcare and other DSA, addressing critical issues such as data privacy, security, access control, and scalability. It proposes solutions to enhance data security and privacy in healthcare systems by examining various existing studies and frameworks. FL is adopted for its ability to facilitate the integration of AI in domains where conven- tional centralized training approaches are infeasible due to data distribution constraints, privacy concerns, or competitive restrictions. Furthermore, FL enables remote or decen- tralized systems to achieve performance comparable to state-of-the-art centralized methods while incurring minimal resource overhead. By eliminating the need for centralized data aggregation, FL significantly mitigates the risks associated with large-scale data breaches and unauthorized access. This thesis makes a substantial contribution in summarizing the existing state of FL, its various types, supporting technologies, and challenges, designing hybrid security frameworks for FL aimed at DSA networks, and performing extensive evaluation of the state-of-the-art FL techniques. Our contributions are as follows: • An extensive systematic literature review has been conducted to investigate the cur- x rent state of FL implementations in DSA fields, such as healthcare and finance. This analysis highlights the complex challenges of implementing FL in mainstream health- care applications, including concerns about transmission costs, data security, privacy, and data/system heterogeneity. Additionally, we provide a detailed taxonomy for the existing literature, focusing on DSA. Alongside identifying the gaps, the review also highlights the fundamental challenges that FL may pose in real-world scenarios. • As organizations and developers explore FL solutions for various applications, the multitude of FL tools and frameworks can feel overwhelming for a beginner. We studied various tools and frameworks available in both open and proprietary forms to provide a concise view of their advantages, limitations, and utility, thereby filling a gap in the proposed literature. • It contributes to improving the security in resource-constrained DSA networks such as IoMT by employing quantization. Unlike prior techniques, we incorporate a workload- aware client selection scheme to overcome the quantization loss, optimize bandwidth, and provide stability to the training network by minimizing the straggler nodes with negligible performance overhead. Thus, providing a reliable security solution to resource-constrained DSA. The experimental results show a 77% to 95% decrease in straggler nodes and an 8% to 0% decrease in accuracy compared to standard FL, depending on the dataset’s complexity. • To ensure security in small edge networks, IoMT networks, and AdHoc networks, a dynamic, lightweight cryptographic FL framework is proposed. Conventional cryp- tographic techniques are computationally expensive for small networks. Thus, we propose a block cipher with an ever-changing key that takes multiple client and net- work attributes into account and is unique to all participants. The overhead in- duced by the cipher and key generation is significantly lower than that of state-of- the-art solutions available. The experimental evaluation indicates that the proposed scheme’s relative performance is comparable to that of native FL. The results reveal an increase of approximately 4.5% to 8% in per-round computation time compared to standard FL. However, this marginal computational overhead is justified by the significantly enhanced security guarantees provided by the proposed approach, par- ticularly in resource-constrained network environments, where conventional security mechanisms with similar computational budgets offer comparatively lower protection • To ensure privacy in networks with sufficient resources, we propose a hybrid FL xi framework that utilizes differential privacy in conjunction with homomorphic encryp- tion. To counter the loss incurred by the differential privacy scheme, we employ the workload-aware client selection. The simulation results demonstrate minimal compu- tational overhead and comparable performance to other state-of-the-art techniques. The proposed Fed-HDVE imposes a minimal time overhead of 2.8 to 3.5 seconds, depending on the model size. The performance evaluation, analysis, and experimental results indicate that the proposed solutions offer a viable and effective security solution. Moreover, the comparative study demonstrates that the suggested approaches perform on par with existing solutions, with lower computational overhead and resource consumption.

FACE DETECTION AND TRACKING

2024-10-01T00:00:00Z

Title: FACE DETECTION AND TRACKING Authors: MOOL, AKSHAY; Panda, Jeebananda (SUPERVISOR); Sharma, Kapil (CO-SUPERVISOR) Abstract: With the use of various technological advancements and devices in today’s routine life, detection and tracking of human faces and facial features become very essential areas of focus. They become important so that more techniques could be developed for increasing their working efficiency. The field of Computer Vision uses these intermediary processes of face detection and tracking to track and analyze the input of visual information about humans, their faces and/or body movements, and correspondingly proceed to the desired application. The present thesis work has been taken up for the development of (i) an optimiz- able face detection and tracking model based on facial landmark localisation and feature tracking, for better and efficient processing of faces in high quality video streams, and (ii) a Non-Neighbourhood Background Elimination component us- ing the built model with mathematical and statistical modelling, for reducing the processing time and computations required for finding the target face in a frame when it has already been detected. Many algorithms have been developed to facilitate Face Detection and Track- ing applications. Viola and Jones were able to develop an algorithm in 2004, that achieves real-time performance with decent accuracy in detecting faces. It was one of the first algorithms to achieve such efficient performance, that is why it is still used as a standard algorithm to compare against other upcoming algorithms, and therefore has been specifically discussed in this thesis. vii There is a lot of visual information that is generated in various fields, rang- ing from daily routine to specialized applications. Processing all these types of information efficiently is a valid concern and need focused research. Most face detection algorithms have to deal with low quality data in videos, since they’re mainly focused on surveillance applications, whose information capturing devices capture less information per frame. Consequently, this thesis reviews some state- of-the-art face detection algorithms and compares their processing efficiency on low and high quality videos. The comparative analysis reveals that these recent and modern algorithms do not work as effectively on high quality videos as they do on lower quality videos. Therefore, there is an increasing need to focus re- search on analysis of high quality information in videos in an efficient manner, so as to keep up the pace of their analysis with the information that is generated. High quality videos (data generated by current applications like social media, Multimedia content, etc) mostly exist in offline mode, that could be used for post processing by the Computer Vision applications. To address this need, an effort has been made to focus on developing such an algorithm that gives faster results on high quality videos, at par with the algorithms working on live low quality video feeds. The proposed algorithm uses Convolutional-MTCNN as base algo- rithm, and speeds it up for high definition videos. The proposed model speeds up the face detection process really fast, up to 19+ FPS, while still maintaining above 90% accuracy. This paper also presents a novel solution to the problem of occlusion and detecting partial or fully hidden faces in the videos. This is achieved by using statistical and probabilistic approaches, given that the face has been identified in first few frames, to give the algorithm an estimate of where the face should be in the occluded region. Since the focus of our research is to efficiently process high quality data, some viii commercially used face detection algorithms in open literature have also been considered in our research. Models like FaceNet, HOG, YuNet, alongwith Viola- Jones algorithm and MTCNN, have been discussed and analysed in our thesis. The research done is compared against these models, in an effort to improve their performance in commercial settings. Further analysis lead to the conclusion that modern face detection algorithms fail to provide optimal results when they have to deal with larger amounts of data per frame while processing higher quality videos. This thesis discusses an- other proposed work that tackles discussed problem and offers a solution to deploy commercially used state-of-the-art face detection algorithms to process only the regions of interest in a frame, and discard the rest to decrease the data to be processed. The model maintains the accuracy of the base algorithm while de- creasing the processing time per frame, thereby increasing the overall efficiency. The selection of region of interest is dependent on the detection of facial window in the previous frame. Therefore, the choice of base algorithm plays an important role in determining the speed of the model. The model achieves increased pro- cessing speeds of about 69–76% more than the standalone usage of the detection algorithms for analyzed frame rates.