4 April 2024 in IEEE Access
Skin lesion classification is a pivotal process in dermatology, enabling the early detection and precise diagnosis of skin diseases, leading to improved patient outcomes. Deep learning has shown great potential for this task by leveraging its ability to learn complex patterns from images. However, diagnostic accuracy is compromised by exclusive reliance on single-modality images. This research work proposes an innovative framework that unifies a Vision Transformer model with transfer learning, channel attention mechanism, and ROI for the accurate detection of skin conditions, including skin cancer. The proposed approach blends computer vision and machine-learning techniques, leveraging a comprehensive dataset comprised of macroscopic dermoscopic images, appended with patient metadata. Compared with conventional techniques, the proposed methodology exhibits significant improvements in various parameters, including sensitivity, specificity, and precision. Moreover, it demonstrates outstanding performance in real-world datasets, reinforcing its potential for clinical implementation. With a remarkable accuracy of 99%, the method outperforms existing approaches. Overall, this investigation underscores the transformative impact of deep learning and multimodal data analysis in the dermoscopic domain, projecting substantial headway into the field of skin lesion analytic diagnosis.
25 June 2024 in IEEE Open Journal of the Computer Society
Whitefly infestations have posed a severe threats to cotton crops in recent years, affecting farmers globally. These little insects consume food on cotton plants, causing leaf damage and lower crop yields. In response to this agricultural dilemma, we developed a novel method for detecting whitefly infestations in cotton fields. To improve pest detection accuracy, we use the combined efficiency of visual transformers and low-cost acoustic sensors. We train the vision transformer with a large dataset of cotton fields with and without whitefly infestations. Our studies yielded encouraging results, with the vision transformer obtaining an amazing 99% accuracy. Surprisingly, this high degree of accuracy is reached after only 10-20 training epochs, outperforming benchmark approaches, which normally give accuracies ranging from 80% to 90%. These outcomes underline the cost-effective potential of the vision transformer in detecting whitefly attacks on cotton crops. Moreover, the successful integration of acoustic sensors and vision transformers opens doors for further research and advancements in the domain of cotton pest detection, promising more robust and efficient solutions for farmers facing the challenges of whitefly infestations.
18 January 2025 in Alexandria Engineering Journal
25 July 2023 in IEEE Sensors Journal
Microbes constitute a substantial portion of Earth’s biodiversity and play a critical role in maintaining the equilibrium of ecosystems. Their significance extends to various domains such as industry, agriculture, and medicine. While most microorganisms pose threats and can cause severe illnesses, a few offer substantial benefits to human life. Traditional methodologies employed in the identification and classification of microorganisms involve time-consuming processes, including staining, biochemical testing, and motility assays. These procedures are indispensable for conducting microbiology experiments. Thus, there is a growing need to integrate deep learning techniques to automate the process of microbial detection. In this study, we propose a novel deep learning model that leverages biosensors to capture real-time data. Extensive analysis is conducted on four diverse datasets collected for this purpose. The empirical results demonstrate that our proposed model outperforms existing methods, achieving an average accuracy exceeding 98% for predicting each microbial class. Notably, this approach also offers practical advantages such as improved feasibility, speed, and cost-effectiveness in microbial detection. By harnessing this state-of-the-art technology, microbiologists can efficiently address complex challenges and obtain optimal results.
9 July 2024 in IEEE MultiMedia
Respiratory illnesses pose a significant threat to life worldwide, necessitating prompt identification and effective intervention. Conventional organ examination methods are restrained by certain shortcomings which lead to inconsistent diagnosis. This research addresses the global threat of respiratory illnesses by introducing a unique Hybrid CNN-GRU architecture deployed on a Raspberry Pi for real-time classification of respiratory auditory cues, leveraging the power of sound in diagnostics. By overcoming the limitations of conventional examination methods, the system was able to achieve an impressive accuracy of 98% in distinguishing unusual auscultations. The system incorporated multimedia elements, particularly sound with CNNs to extract spatial attributes and GRUs for the comprehension of temporal context. The utilization of an instinctual online interface complemented with visualizations, dynamic sound patterns, and interactive elements eased direct communication with medical professionals. The multimedia-centred approach particularly on respiratory sound indicates a landmark of respiratory diagnostics that is poised to enhance healthcare outcomes globally.