Detecting kidney stones in coronal CT images remains challenging due to the small size of stones, anatomical complexity, and noise from surrounding objects. To address these challenges, we propose a deep learning architecture that augments a Vision Transformer (ViT) with a pre-processing module. This module integrates CSPDarknet for efficient feature extraction, a Feature Pyramid Network (FPN), and Path Aggregation Network (PANet) for multi-scale context aggregation, along with convolutional layers for spatial refinement. Together, these trained components filter irrelevant background regions and highlight kidney-specific features before classification by ViT, thereby improving accuracy and efficiency. This design leverages ViT’s global context modeling while mitigating its sensitivity to irrelevant regions and limited data. The proposed model was evaluated on two coronal CT datasets (one public and one private dataset) comprising 6,532 images under six experimental scenarios with varying training and testing conditions. It achieved 99.3% accuracy, 98.7% F1-score, and 99.4% mAP@0.5, higher than both YOLOv10 and the baseline ViT. The model contains 61.2 million parameters and has a computational cost of 37.3 GFLOPs, striking a balance between ViT (86.0M, 17.6 GFLOPs) and YOLOv10 (22.4M, 92.0GFLOPs). Despite having more parameters than YOLOv10, the model achieved a lower inference time than YOLOv10, approximately 0.06 seconds per image on an NVIDIA RTX 3060 GPU. These findings suggest the potential of our approach as a foundation for clinical decision-support tools, pending further validation on heterogeneous and challenging clinical datasets such as small (<2 mm) or low-contrast stones.

The rise of virtual instruments has revolutionized music production, providing new avenues for creating music without the need for physical instruments. However, these systems rely on costly hardware, such as MIDI controllers, limiting accessibility. As an alternative, 3D gesture-based virtual instruments have been explored to emulate the immersive experience of MIDI controllers. Yet, these approaches introduce accessibility challenges by requiring specialized hardware, such as depth-sensing cameras and motion sensors. In contrast, 2D gesture systems using RGB cameras are more affordable but often lack extended functionalities. To address these challenges, this study presents a 2D virtual piano system that utilizes hand gesture recognition. The system enables accurate gesture-based control, real-time volume adjustments, control over multiple octaves and instruments, and automatic sheet music generation. OpenCV, an open-source computer vision library, and Google’s MediaPipe are employed for real-time hand tracking. The extracted hand landmark coordinates are normalized based on the wrist and scaled for consistent performance across various RGB camera setups. A bidirectional long short-term memory (Bi-LSTM) network is used to evaluate the approach. Experimental results show 95% accuracy on a public Kaggle dynamic gesture dataset and 97% on a custom-designed dataset for virtual piano gestures. Future work will focus on integrating the system with Digital Audio Workstations (DAWs), adding advanced musical features, and improving scalability for multiple-player use.

This study addresses the challenge faced by tourists, companies, travel agents, and tourism agencies in selecting the ideal hotel in Samarinda, given the variety of available options. The city boasts numerous hotels with differing facilities, room types, rates, and locations, which can complicate decision-making without adequate information. To provide a solution, this research introduces a Decision Support System (DSS) that employs the Multi-Attribute Utility Theory (MAUT) method for hotel assessment. By evaluating hotels based on key attributes like price, amenities, service quality, and location, the system offers a comprehensive, objective approach to determining the best affordable hotels. The study contributes significantly to the hospitality sector by presenting a practical tool that simplifies the hotel selection process and ensures that choices align with the preferences of the visitors.

This research work aims to utilize deep learning techniques to identify autism traits in children based on their facial features. By combining traditional convolutional approaches with attention layers, the study seeks to enhance interpretability and accuracy in identifying autism spectrum disorder (ASD) traits. The dataset includes diverse facial images of children diagnosed with ASD and neuro-typical children, ensuring comprehensive representation. Preprocessing techniques standardize and enhance image quality, mitigating biases. Integration of attention layers within the convolutional neural net-work (CNN) architecture focuses on crucial facial features, improving feature extraction and classification accuracy. This approach enhances model interpretability through eXplainable AI (XAI) techniques. Model training involves optimization and validation processes, employing hyper parameter tuning and cross-validation for robustness. The performance of this combined model yielded close to 95% accuracy outperforming existing models in terms of complexity to accuracy ratio.

In the healthcare field, the detection of critical diseases such as brain tumors is essential. A technique like traditional support vector machine has been commonly used for brain tumor classification. However, Processing and detecting brain tumors requires achieving high accuracy with shorter detection time and reduced complexity. To accomplish this, efficient feature selection is necessary, which can be based on various factors. A convolutional neural network-based stacking technique is introduced for effective brain tumor classification and prediction using Red Panda optimization. By efficiently extracting spatial data from medical images, a convolutional neural network is used in stacking to enhance thecapacity of our model for abnormality detection and classification in the prediction of brain tumors. Red panda optimization is a biologically inspired stochastic optimization algorithm used for the effective selection of significant features. This Technique improves the prediction accuracy in a shorter period and reduces the complexity by selecting significant features for a huge amount of data by employing effective optimization. This technique is tested on multiple standard datasets to assess our model’s performance. Our technique is compared to other optimization models such as Mutual information-based optimization and traditional particle swarm optimization for further validation. Our model showed an improvement in detection accuracy to 98% with a better reduction in detection time and complexity.

To augment the accuracy of the results of a Time-Series Forecasting problem in the Computational Epidemiology domain of Public Health, to generate an accurate alert in a Real-time Outbreak and Disease Surveillance (RODS) system, namely in the prediction of Malaria incidences, an interdisciplinary approach of data analysis [through Statistical along with Machine-Learning (ML) and Deep-Learning techniques (DL)] has been studied in this research. Two different Non-linear Deep-Learning based techniques, viz., Long Short-Term Memory (LSTM) [a subclass of Recurrent Neural Network (RNN)] & Gated Recurrent Unit (GRU) and two different Non-linear Machine-Learning techniques, viz., Random Forest Regressor & Non-linear Support Vector Machine Regressor are applied in this study to compare against the traditional Statistical-based linear SARIMA model, to forecast a longitudinal data-set of malaria incidences. While SARIMA or other traditional Autoregressive (AR) models, necessitating a smaller number of parameters, undergo limited training and limited prediction power, ML and DL models show profound and persistent performance improvement with better noise-handling/ missing values and perform multi-step forecasts. Moreover, the over-fitting issue can be combated by introducing densely connected residual links in the ML/ DL networks.

Efficient comparison of heterogeneous tabular datasets is difficult when sources are unknown or weakly documented. We address this problem by introducing a unified, type-aware framework that builds compact data represen- tations (CDRs)—concise summaries sufficient for downstream analysis—and a corresponding similarity graph (and tree) over a data corpus. Our novelty is threefold: (i) a principled vocabulary and procedure for constructing CDRs per variable type (factor, time, numeric, string), (ii) a weighted, type-specific similarity metric we call Data Information Structural Similarity (DISS) that aggregates distances across heterogeneous summaries, and (iii) an end-to-end, cloud-scalable real- ization that supports large corpora. Methodologically, factor variables are summarized by frequency tables; time variables by fixed-bin histograms; numeric variables by moment vectors (up to the fourth order); and string variables by TF–IDF vectors. Pairwise similarities use Hellinger, Wasserstein (p=1), total variation, and L1/L2 distances, with MAE/MAPE for numeric summaries; the DISS score combines these via learned or user-set weights to form an adjacency graph whose minimum-spanning tree yields a similarity tree. In experiments on multi-source CSVs, the approach enables accurate retrieval of closest datasets and robust corpus-level structuring while reducing storage and I/O. This contributes a repro- ducible pathway from raw tables to a similarity tree, clarifying terminology and providing algorithms that practitioners can deploy at scale.