Layer-Wise Fine-Tuning of ResNet Models for Enhanced Skin Cancer Diagnosis with Dermoscopic Image Sakshi Matiray, Shashi Bhushan 2025 7th International Conference on Information Systems and Computer Networks Iscon 2025, 2025 Environmental and genetic factors have made skin cancer, especially melanoma, an increasing global health concern. While early detection is essential, conventional techniques can be laborious and call for specialized knowledge. Utilizing the ISIC 2019 dataset, this work investigates the efficacy of deep learning models, more specifically, ResNet18, ResNet50, and ResNet152, for automated skin lesion categorization. Some experimentation techniques have been used, such as partial layer training, finetuning, and complete training. Accuracy, precision, and recall are also implemented to assess the models. After fine-tuning the final 5% of layers, ResNet50 outperformed the other evaluated designs, achieving an ROC AUC of 0.962. The study shows that deep networks can be somewhat fine-tuned to save training time without sacrificing classification accuracy. These results open the door for further integration with clinical metadata and implementation in practical diagnostic tools, confirming the potential of AI-based approaches for early and scalable skin cancer detection.
AI-Based Model for Predicting and Managing Dual - Peak Electricity Demand in Delhi Shashi Bhushan, Ramya Sharma, Harsh Mishra, Abhiraj Krishna Babu Proceedings of the 2025 International Conference on Technology Enabled Economic Changes Intech 2025, 2025 The surveys show that the electricity demand pattern of NCT of Delhi is not only a peak in summer but twice in a day, once in the morning and second in the evening due to odd demand. There is a duck shaped curve due to the integration of solar energy. Thirdly, it is demand for management is also an area affected by uneven growth of the urban environments. Therefore, the present research presents an Artificial Intelligence based model for forecasting as well as for controlling the electricity usage of the city Delhi. The model also involves responding to other conditions such as weather conditions, holidays, load development patterns, and real estates in order to solve issues of load restructure as well as the power purchasing schemes. Thus, the following benefits can be seen for the model: The efficiency of the distributed resources' utilization; The effectiveness of the grid; Effective planning of energy sustainably. As was expected earlier, the levels of accuracy have remained high hence the output is very valuable to stakeholders in the energy sector in Delhi. It includes solution information pertinent to future energy problems that would assist the stakeholder in addressing them.
AI-Based web application firewall for real-time malicious query detection Shashi Bhushan, Harsh Mishra, Ramya Sharma, Abhiraj Krishna Babu, Harshit Verma Recent Trends in Intelligent Computing and Communication Volume 1, 2025 This research paper explores the development and implementation of an AI-driven Web Application Firewall (AI-WAF) designed to identify and mitigate malicious web queries effectively. By using deep learning, the AI-WAF has achieved high performance, such as the accuracy and balanced F1 scores over 98% and above 97%, respectively. The paper shows the advantage of artificial intelligence-based (AI-based) methods over conventional rule-based methods, highlighting their flexibility, scalability, and robustness in terms of real-time threat detection. The paper also makes a comparison between the AI-WAF and current models, in this way showing its superior performance in characterizing new attack patterns. Principal contributions are represented by novel methods for the strengthening of the security of web applications, and by suggestions for the strengthening of scalability, dataset heterogeneity, and complexity of hybrid defence systems. AI-WAF's potential future uses in IoT security, as well as integration with threat intelligence systems, are highlighted.
Performance Benchmarking of Deep Learning Techniques for Classifying Thyroid Tumors in Ultrasound Scans Shashi Bhushan, Harsh Mishra, Ramya Sharma, Abhiraj Krishna Babu, Harshit Verma 2025 7th International Conference on Information Systems and Computer Networks Iscon 2025, 2025 One of the most important aspects of global health today is the concern for cancer, including thyroid cancer, which needs special attention related to early detection and diagnosis in order to treat it effectively. This paper focuses on classifying thyroid tumors using deep learning techniques and proposes a custom-designed convolutional neural network (CNN) architecture specifically optimized for ultrasound images. CNN is used to classify thyroid nodules as benign or malignant. The data set consists of ultrasound images which are publicly available. Several image processing approaches are performed in order to improve model performance. Diagnostic evaluation of the proposed model is performed using classification metrics such as accuracy, precision, recall, F1 score, and AUC-ROC curve analysis. The findings indicate that deep learning approaches are able to accurately classify different types of thyroid tumors as well as surpass the performance of conventional machine learning methods. Moreover, the study displays visual analyses such as the accuracy/loss curve, confusion matrix plot, and ROC-AUC curve plots to show the model performance more comprehensively. Explainability is further enhanced through Grad-CAM and SHAP visualizations, aiding in clinical interpretation.
Deep Learning Approach for Hyper-Multiclass Consumer Electronics Image Clustering Using Contrastive Learning Ajmeera Kiran, Janjhyam Venkata Naga Ramesh, Vrince Vimal, Kishore M. Kumar, Mukesh Soni, Shashi Bhushan, Tariq Ahamed Ahanger, Pavitar Parkash Singh, Rajesh Singh IEEE Transactions on Consumer Electronics, 2024 Efficient management of image data is essential in the consumer electronics industry. Image clustering has evolved through the utilization of dimensionality reduction and representation learning techniques to extract pertinent features. However, dealing with a broad spectrum of Consumer Electronics Image categories poses challenges due to complex data distribution and cluster density. To tackle these issues, a contrastive learning-based hyper-multiclass deep image clustering approach has been introduced. This model comprises three main steps. Firstly, to accommodate the diverse range of consumer electronics images, the feature model undergoes initial training using enhanced contrastive learning to ensure uniform cluster distribution. Secondly, semantic similarity is leveraged to acquire instance semantic closest neighbor data employing a multi-perspective methodology. Lastly, instances along with their closest neighbors serve as self-supervised information to train the clustering model. Ablation and comparative analyses have demonstrated the efficacy of this approach. It effectively achieves uniform distribution of clusters within the mapped space and consistently extracts semantic nearest neighbor information. Comparative experiments were conducted on benchmark datasets, specifically the ImageNet G200 and G1000 class datasets, within the consumer electronics context. Results indicate a significant improvement in accuracy, with enhancements of 10.6% and 9.2%, respectively, surpassing the performance of existing advanced methods.
