Modigari Narendra

Introduction: Breast cancer diagnosis using histopathological imaging remains a critical yet challenging task, requiring automated systems that generalize reliably across patients and varying imaging conditions. While deep learning models have shown strong performance, many prior studies employ image-wise data splits that introduce patient-level data leakage, resulting in overly optimistic and potentially misleading evaluations. This study aims to address this limitation by establishing a rigorous, leakage-free benchmarking framework for binary breast cancer histopathology classification. Methods: A comprehensive evaluation of nine deep learning architectures was conducted on the BreaKHis dataset, comprising 7,909 images from 82 patients. The models include six convolutional neural networks (ResNet50, MobileNetV2, VGG16, DenseNet121, Xception, and EfficientNetB0), one modern convolutional architecture (ConvNeXt), and two transformer-based models (Swin-Small and Swin-Base). A strict 5-fold patient-aware cross-validation protocol was implemented to ensure that images from the same patient were not shared between training and validation sets. All models were trained under identical experimental conditions. Performance was assessed using accuracy, precision, recall, and F1-score, reported as mean ± standard deviation. Statistical significance was evaluated using paired t -tests and Wilcoxon signed-rank tests with Bonferroni correction. Results: All evaluated architectures demonstrated comparable performance, achieving mean accuracies in the range of 0.91-0.93. ResNet50 achieved the highest mean accuracy (0.9267 ± 0.0435) and F1-score (0.9472), although differences among models were marginal. Statistical analysis confirmed that no pairwise differences were statistically significant ( p > 0.05 after correction). Magnification-wise analysis indicated that intermediate resolutions (40× and 200×) provided more discriminative features, whereas higher magnification (400×) resulted in reduced performance due to limited contextual information. Discussion: The findings highlight that, under a rigorously controlled and leakage-free evaluation protocol, architectural differences among modern deep learning models do not lead to statistically significant performance variations. Instead, evaluation design plays a more critical role in determining reliable outcomes. The proposed patient-aware benchmarking framework enhances reproducibility and provides a robust foundation for future research, supporting the development of clinically translatable AI systems for breast cancer diagnosis.

Introduction: Substance use continues to evolve as a multidimensional public health challenge influenced by traditional behavioral triggers and emerging digital interactions. This study investigates how demographic factors, psychological states, and patterns of digital engagement shape substance use behaviors using multiple behavioral data sources. Methods: Quantitative analyses were conducted using the NHANES dataset and a Kaggle social media psychology dataset to identify statistical relationships and train predictive machine learning models for substance use indicators and digital behavioral patterns. Random Forest, XGBoost, AdaBoost, Support Vector Regression (SVR), and Logistic Regression models were evaluated, with hyperparameter tuning applied to improve predictive performance. In addition, a supplementary survey ( N = 236) was collected and used as a qualitative interpretive layer to contextualize the relationship between digital behavior and substance use risk. Results: The analysis revealed nonlinear relationships between social media engagement, anxiety, and loneliness. Contrary to the widely cited linear dose–response assumption, anxiety scores plateaued at higher levels of digital engagement, suggesting that the qualitative nature of online interactions may exert greater influence on psychological distress than usage duration alone. Machine learning models demonstrated improved predictive performance after hyperparameter tuning across both datasets. Discussion: These findings highlight the importance of considering digital engagement patterns alongside traditional behavioral and demographic factors in substance use research. The results support the development of platform-specific digital well-being strategies, nuanced behavioral modeling approaches, and culturally sensitive interventions that integrate both objective behavioral data and subjective user experiences. The proposed multi-source evidence framework provides a foundation for future exploratory behavioral risk profiling and prevention systems.

Real-time object detection in unmanned aerial vehicle based Search and Rescue missions requires a critical balance between accuracy, speed, and the computational constraints of edge devices. YOLOv10 provides a powerful baseline, but operational efficiency greatly depends on the backbone architecture and its performance. In this study, an evidence-based framework is set, which would help search and rescue practitioners to pick the most appropriate model configurations according to their specific mission profiles and hardware capabilities.We conducted a fully empirical study to understand how to select a model from 8 different backbones, including different versions of EfficientNet, GhostNet, MobileNet, ResNet50, and Swin, across the five YOLOv10 variants from Nano to X-Large on the Search and Rescue Dataset. The results map a clear performance landscape where models are categorized into distinct groups based on the trade-off between inference speed, accuracy, and computational load. For the edge deployment with high constraints, the YOLOv10-Nano with MobileNet backbone achieves the best compromise, at which it exhibits a very high inference speed of 80.5 FPS with 0.3945 mAP@0.5:0.95 accuracy. For server environments with low or no constraints, YOLOv10-XLarge with EfficientNetV2M attained the highest performance of 0.4898 mAP with 29.3 FPS. Yolov-medium with the EfficientNetV2Mbackbone achieved a very good tradeoff between accuracy and speed, with a map@50:95 score of 0.4791 with close to 54 frames per second. Our work highlightshts new light on the trade-offs between speed, accuracy, and complexity of models for more effective and resource-aware object detection systems, thereby speeding up rescue operations.

Introduction: Electrocardiograms (ECGs) play a crucial role in diagnosing heart diseases by capturing the electrical activity of the heart. With the rising need for real-time cardiac monitoring, portable solutions have gained significance for timely detection and intervention. This study presents a portable ECG monitoring system incorporating Convolutional Neural Networks (CNNs) for accurate classification of cardiac abnormalities, including arrhythmias.Methods: The proposed system consists of an Arduino Nano microcontroller interfaced with an AD8232 ECG sensor for real-time ECG signal acquisition. The collected ECG data undergoes preprocessing before being fed into CNN models trained on the MIT-BIH Arrhythmia dataset. The model is designed for both binary and multi-class classification, distinguishing normal and abnormal heart rhythms. Performance metrics, including accuracy, were evaluated against state-of-the-art approaches to assess classification effectiveness.Results: Experimental evaluations demonstrate the CNN model’s high classification accuracy, achieving 98.35% in binary classification and 99.3% in multi-class classification. These results surpass existing benchmarks, highlighting the efficiency of the proposed system. The system's low-cost hardware and real-time classification capabilities enhance its suitability for continuous cardiac monitoring.Discussion: The proposed ECG monitoring system presents a reliable and cost-effective solution for early arrhythmia detection. By leveraging CNNs, it ensures accurate classification of cardiac abnormalities, making it a promising tool for both clinical and remote healthcare settings. Its potential impact extends to real-time monitoring, early diagnosis, and personalized healthcare, contributing to improved cardiovascular health management.

The human skin serves as a critical indicator of underlying health conditions, often manifesting early signs of disorders affecting internal organs. Recognizing these signs is crucial for timely diagnosis and treatment. However, the significance of the body’s natural defence mechanism through the skin is sometimes overlooked. This study aims to develop a skin disease classification system using a multimodal approach, integrating ensemble methodology using DenseNet169 – Resnet50 ensemble Transfer learning models, and Natural Language Processing (NLP) within a Telegram chatbot interface. The primary aim is to enhance the chatbot’s ability to deliver customised skin-related diagnoses by using user-provided information, including skin type, chemical exposure, and previous treatments, ensuring more precise and personalised interactions. Additionally, the accuracy and generalization capability of the classification system is improved by analysing the data from both the chatbot and image analysis. The chatbot’s self-learning capabilities allow it to improve its comprehension over time in response to user input, which makes it more adept at personalising queries. Integrating DenseNet169 and ResNet50 strengthens the feature reuse by connecting each layer to every other layer in a dense manner for efficient gradient flow thus reducing the number of parameters. ResNet50, with its residual connections, helps mitigate vanishing gradient issues, allowing for deeper networks with stable training leading to improved feature extraction and representation. This hybrid approach captures fine-grained global features enhancing the feature learning attaining good performance in complex tasks. A total of 11,747 images were evaluated in the proposed study, with 7,930 images allocated for training and validation, and 3,817 images designated for testing. The model attained an accuracy of 77.07% and an AUC score of 96.72%. The NLP training model achieved an accuracy of 93.62% ensuring a comprehensive understanding of user data and accelerating engagement, leading to more precise and personalized predictions of skin disorders.

In renal disease classification, various pathologies such as kidney stones, renal cysts, and tumors pose significant health risks and diagnostic challenges. Annually, thousands of cases are diagnosed worldwide, impacting both pediatric and adult populations. To address this, Computer Assisted Diagnosis (CAD) systems have emerged, aiming to alleviate the burden on healthcare professionals, enhance diagnostic accuracy, and manage the vast amounts of medical data involved. These CAD systems traditionally leverage the expertise of nephrologists, specialized imaging features, and domain knowledge to aid in disease classification. The MSKd_Net architecture integrates Swin Transformer-based hierarchical learning, efficiently capturing both local and global features, along with a customized convolved model enhanced with a multi-head attention layer. Additionally, the custom convolved model enhanced with the multi-head attention layer efficiently captures long-range dependencies and relationships in the input sequence or set of features, resulting in fine-tuned features. The Swin Transformer features are fused additively with the convolved model enhanced with multi-head attention layer features and fed to linear layers to classify the input such as cyst, stone, tumor, and normal. The proposed method significantly enhanced the discrimination capability of the MSKd_Net model on the Riñones dataset, achieving optimal F1-scores of 0.97, 0.99, 0.98, and 0.9 for cyst, normal, stone, and tumor, respectively on testing data. The strategic fusion of these components effectively addresses classification challenges, enhancing crucial feature identification for accurate multi-class kidney disease image recognition, and seamlessly integrates within MSKd_Net to achieve superior performance.