Developed a novel BMFCNet architecture integrating blended multi-level feature extraction for robust detection of Major Depressive Disorder (MDD) from EEG signals. Introduced a Residual-Inception module to effectively capture both low-level (LL) and high-level (HL) discriminative EEG features, enhancing representational capacity. Developed a Constraint Fusion mechanism for adaptive weighting and fusion of LL and HL features, improving feature integration and classification performance. Addressed subjectivity in MDD diagnosis by providing an automated, EEG-based framework that enhances accuracy, reliability, and clinical applicability. Validated the proposed model on benchmark datasets, demonstrating superior performance compared to 16 state-of-the-art methods in terms of accuracy and efficiency.

TinyEyeNet introduces a lightweight CNN for accurate anterior segment eye disease classification Designed for high performance with low computational cost, ideal for real-world deployment Trained and validated on a custom-curated clinical eye image dataset Achieves strong diagnostic accuracy, outperforming conventional deep models Suitable for resource-constrained and portable ophthalmic screening systems Enables faster, scalable, and accessible eye disease detection

• Introduces the RAIE Transformer, a lightweight, recency-augmented architecture designed specifically for accurate cricket ball trajectory prediction using short input sequences. • Uses a curated dataset of 4,400 frames from 212 cricket deliveries, manually annotated and preprocessed for reliable spatio-temporal modelling. • Incorporates a novel Recency-Augmented Input Embedding (RAIE) that prioritizes recent motion cues while preserving long-term context to improve prediction accuracy. • Outperforms traditional models such as LSTM, RNN, Base Transformer, Informer, TimeMixer, and Graph Transformer—achieving MSE 3.95, ADE 2.35, and FDE 2.56, the best among all tested models. • Provides a fully automated end-to-end pipeline combining YOLO-based detection, interpolation, and autoregressive prediction—eliminating manual annotation requirements. • Demonstrates strong real-time potential, handling non-linear motion (swing, spin, bounce) with high efficiency and low computational cost, suitable for sports analytics and broadcast applications

• Proposes an AI-assisted otoscope framework for automated ear disease diagnosis • Enables early and accurate detection of common ear conditions from otoscopic images • Utilizes a deep learning–based image analysis pipeline for robust classification • Reduces dependency on specialist expertise, supporting primary healthcare screening • Designed for real-time and portable diagnostic applications • Demonstrates high classification accuracy on clinically relevant datasets • Enhances accessible, scalable, and cost-effective ear healthcare solutions.

• Proposes a hybrid activation function (ReLU + ELU) to overcome limitations of existing functions. • Addresses issues like dying ReLU and vanishing gradients. • Demonstrates improved performance on CNN and MLP models using MNIST & FashionMNIST datasets. • Shows better convergence and learning efficiency compared to standard activation functions.

• Proposes SARANet, a self-attentive multi-scale residual attention network for Braille recognition. • Combines R-CBAM (channel & spatial attention) and Compact Self-Attention (CSA). • Captures fine-grained details and long-range dependencies in Braille images. • Works on Arabic Braille dataset with 32 character classes. • Achieves 95.29% accuracy and F1-score of 0.952. • Improves optical Braille recognition (OBR) performance. • Supports assistive technology for visually impaired individuals.

Developed SpineDeep-Net, a novel deep learning framework integrating self-attention with multi-layer CNNs for automated selection of diagnostically relevant lumbar MRI disc slices. Eliminated manual slice selection by automatically identifying optimal transverse mid-height slices of L3/L4, L4/L5, and L5/S1 IVDs, improving efficiency and reducing human error. Introduced a dual self-attention mechanism, enabling enhanced long-range spatial dependency modeling and precise localization of subtle intervertebral disc variations. Enhanced feature representation through dynamic feature recalibration and strengthened inter-feature dependencies, improving discriminative learning for lumbar spine analysis. Demonstrated superior performance on experimental evaluation, achieving 96.83% accuracy and 98.41% specificity, outperforming state-of-the-art methods and supporting scalable clinical deployment.

• Proposes GAT-TNBC-Net, a dual-branch attention-based model for TNBC subtype classification. • Combines Lightweight Attentive Head Network (LAHN) and Graph Attention Network (GAT). • Captures both global gene dependencies and structural gene relationships. • Uses multi-head attention fusion for integrating tabular and graph features. • Works on high-dimensional RNA-seq data (55,662 genes per sample). • Achieves 97.22% test accuracy and 92.22% cross-validation accuracy. • Supports personalized treatment planning with interpretable predictions.

• Lightweight multi-attention CNN designed for real-time knot classification. • Achieves 95.16% test accuracy, outperforming all benchmark models. • Uses Ghost, SE, and CBAM modules with stochastic depth for efficiency. • Computationally efficient at 1.03 GFLOPs with 59 ms inference time. • Grad-CAM explainability highlights knot-specific visual cues for trust and transparency. • Robust performance across lighting, background, and rope tension variations.

Developed RGTFormer, a novel hybrid deep learning architecture integrating a categorical gated transformer with a Relational Graph Convolutional Network (RGCN) for mutation-driven drug resistance prediction in TB. Leveraged multi-modal mutation features, incorporating both genomic sequence and structural characteristics from six key drug-resistance genes to enhance predictive capability. Modeled inter-mutation dependencies using RGCN, enabling effective capture of relational information among mutations, while the transformer captures complex feature interactions. Demonstrated superior predictive performance, achieving 98.67% test accuracy and 97.15% cross-validation accuracy, outperforming conventional ML and DL baselines. Validated model effectiveness through ablation studies, highlighting the contribution of gated attention and graph-based learning, and providing an interpretable framework for clinically relevant resistance prediction and personalized TB treatment.

• Next-Gen AI Safety: SalienNext uses an upgraded ConvNeXt + CBAM attention design to spot drowsy drivers with far higher accuracy than traditional vision models. • Ultra-Sharp Eye Detection: Multi-stage attention zooms into subtle eye and facial cues, enabling the system to detect even minor signs of fatigue. • State-of-the-Art Accuracy: Achieves an impressive 99% accuracy on the MRL Eye Dataset, outperforming leading deep learning models in the field. • Lightweight & Real-Time Ready: Optimized for speed and efficiency, making it deployable on edge devices inside vehicles. • Smart Attention Fusion: Combines low-level details and high-level semantics for a clearer understanding of drowsiness signals. • Designed for Safer Roads: Built to support next-gen ADAS and intelligent transportation systems, targeting reduced accidents and enhanced driver monitoring.

• MultiFlipFormer is a multimodal transformer designed for emotion flip reasoning and instigator detection in therapeutic conversations. • It models dynamic emotional transitions across dialogue turns using textual, visual, and contextual cues. • Trained on the MESC dataset with over 28,000 utterances covering 7 emotions and 10 therapy strategies. • Achieves a weighted F1-score of 0.828 and perfect instigator detection accuracy (1.000). • Outperforms existing models like TGIF and MPT-HCL, offering real-time therapeutic insights and emotion trajectory forecasting. • Future work focuses on explainability, few-shot learning, and multilingual adaptation for broader clinical use.

• Introduces a resource-efficient reinforcement learning framework using Group Relative Policy Optimization (GRPO) to enhance mathematical reasoning in small-scale language models. • Implements structured prompting with XML-style tags to clearly separate reasoning steps from final answers, improving interpretability and training alignment. • Integrates memory-efficient optimization techniques—including 8-bit AdamW, mixed precision, gradient checkpointing, and accelerated decoding—to enable single-GPU fine-tuning. • Achieves 50.95% accuracy on the GSM8K benchmark, outperforming larger models despite using only a 0.5B parameter model. • Demonstrates that compact LLMs can achieve competitive reasoning ability, making reinforcement learning viable in low-resource environments.

• Proposes a Self-Attention U-Net integrated with Residual Bottleneck layers to generate high-quality MRI images from low-cost CT scans. • Utilizes self-attention mechanisms to capture long-range dependencies, enhancing anatomical and structural accuracy in medical image translation. • Incorporates a spectral-normalized Patch Discriminator to ensure realistic MRI-like outputs with improved perceptual quality. • Achieves strong results with PSNR up to 25.78 dB and SSIM up to 0.7125, outperforming baseline models like U-Net, Pix2Pix, and CycleGAN. • Offers a cost-effective solution for medical imaging in resource-limited regions, improving diagnostic accessibility and healthcare outcomes.

• Introduces a lightweight image splicing detection framework based on information-theoretic principles, avoiding the need for deep learning or GPUs. • Extracts 20 handcrafted features across spatial, cross-channel, and multi-scale domains using entropy, mutual information, and Kolmogorov complexity. • Achieves 85.32% accuracy, 0.934 AUC-ROC, and 0.8571 F1-score on the Columbia Image Splicing Dataset, rivaling deep models. • Provides interpretable results through feature importance analysis, identifying key tampering indicators like information variance and conditional entropy. • Demonstrates a computationally efficient, explainable, and generalizable solution for digital image forgery detection in forensic and media applications.