Research
Machine learning research in healthcare AI. medical imaging, NLP for mental health, and deep learning systems.
2025
ACB-TriNet: Asymmetric Convolutions and Triplet Attention for Effective Malware Classification
International Conference on Emerging Trends in Cybersecurity (ICETCS 2025, UK)
Springer
🏆 Best Technical Paper
Accepted
With malware's rapid expansion and increasing complexity, classic detection methods based on signatures and behavioral analysis have significant limitations. This study introduces ACB-TriNet, a dual-branch deep learning architecture that converts malware samples into three-channel image representations using grayscale, entropy, and Sobel edge features. These representations enable the model to capture both global structural patterns and finely detailed local textures. The framework includes Asymmetric Convolution Blocks (ACBs) for directional feature extraction, Triplet Attention for cross-dimensional refinement, and a Global Attention Block (GAB) for final feature fusion. A class-balanced focal loss is used to reduce data imbalance and increase minority-class sensitivity. Experiments with the Malimg dataset demonstrate that ACB-TriNet achieves 98.98% accuracy, a 98.81% F1-score, and a 2.28% false negative rate, drastically reducing misclassification errors and exceeding previous attention-based models.
2024
Early Detection of Suicidal Ideation Using Bidirectional GRU and Language Models
3rd International Conference on Computing Advancements (ICCA)
ACM
Suicide has recently emerged as a leading cause of death worldwide, underlining the importance of effective preventative measures. Online social media posts can provide valuable insights into people who are suicidal and assist in preventing unfortunate outcomes. This study examined the utilization of Bidirectional GRU to improve text classification using language models, incorporating Bi-GRU layers with popular pre-trained language models like BERT, RoBERTa, DistilBERT, DistilRoBERTa, and ELECTRA-Small. The BERT-BiGRU and DistilBERT-BiGRU models demonstrated notable effectiveness, achieving accuracies of 95.8% and 95.2% respectively, with remarkably low false negative rates of 4.17% and 2.80%.
Improving Pre-Trained CNNs with CBAM and Skip Connections for Multi-Class Retinal Diseases Classification using OCT Images
3rd International Conference on Computing Advancements (ICCA)
ACM
Millions of people suffer from retinal defects worldwide. Early discovery and treatment of these anomalies could halt further progression, saving many people from preventable blindness. This study presented a hybrid framework utilizing pre-trained models (DenseNet121, ResNet50, VGG16, Xception, and EfficientB1) incorporating the Convolutional Block Attention Module (CBAM) and skip connections for accurate retinal disease classification. The DenseNet-CBAM-Skip and Xception-CBAM-Skip architectures achieved high accuracies of 96.28% and 96.11% respectively.
Advancing Glioma Segmentation: A Robust 3D Residual Attention U-Net Framework for Multimodal MRI Images
3rd International Conference on Computing Advancements (ICCA)
ACM
Brain tumors are abnormal growths of cells within the brain, posing significant health challenges. Glioma, originating from supportive glial tissue, is notably concerning due to its low survival rate. This study presents a 3D Residual Attention U-Net architecture that integrates spatial and channel attention mechanisms for enhanced feature representation. Using a modified Focal-Dice loss function to handle class imbalance, the proposed architecture achieves a Dice coefficient of 0.9002 and an IoU of 0.8272.
2023
Multi-class Brain Tumor Classification with DenseNet-Based Deep Learning Features and Ensemble of Machine Learning Approaches
2nd International Conference on Big Data, IoT and Machine Learning (BIM)
Springer
This study proposed a two-phase end-to-end framework comprising DenseNet-121-based deep learning for feature extraction and an ensemble of machine learning methodologies for precise brain tumor classification. Preprocessing MRI images to eliminate unwanted regions enhanced the deep learning model's feature extraction capabilities. The ensemble mechanism achieved an accuracy of 98.86% and an F1-score of 98.76% without any data augmentation, with random forest attaining the highest individual performance.
A Late Fusion Deep CNN Model for the Classification of Brain Tumors from Multi-Parametric MRI Images
International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM)
IEEE
This study proposed a late fusion CNN architecture that integrates features extracted from each MRI sequence at a later stage in the classification process, allowing the model to capture unique features of each sequence while leveraging complementary information. The implementation of this cutting-edge deep learning-based late fusion multi-parametric brain tumor classification approach achieved 97% test accuracy.
Attention Mechanism-Enhanced Deep CNN Architecture for Precise Multi-class Leukemia Classification
2nd International Conference on Big Data, IoT and Machine Learning (BIM)
Springer
Our proposed deep learning architecture combines transfer learning with attention mechanisms to classify subtypes of leukemia accurately. Using a publicly available dataset of blood cell images adhering to WHO standards, our DenseNet201 with CBAM model achieves a remarkable 99.85% overall accuracy without resorting to data augmentation, surpassing previous methods and attaining state-of-the-art results in leukemia classification literature.