A Review of Explainable Federated Learning Frameworks for Chest X-ray Diagnosis under Heterogeneous Hospital Data
The application of deep learning in chest X-ray diagnosis has demonstrated promising results in detecting multiple thoracic diseases. However, traditional centralized approaches face significant challenges, including limited generalization across hospitals with heterogeneous patient populations and imaging protocols, compounded by strict privacy regulations such as the Health Insurance Portability and Accountability Act (HIPAA) and the General Data Protection Regulation (GDPR) that prevent data sharing between institutions. Although centralized deep learning approaches perform well and achieve high local accuracy on their predictions, they are often “black boxes,” which limits clinical trust and interpretability. This review examines existing explainable federated learning frameworks for chest X-ray diagnosis under heterogeneous data conditions. Current approaches enable decentralized training across non-independent and identically distributed (non-IID) hospital environments, utilizing robust aggregation strategies such as Federated Averaging (FedAvg) and Federated Proximal (FedProx) to address label, quantity, and feature skew. To establish clinical trust, explainable Artificial Intelligence (XAI) techniques, such as Gradient weighted Class Activation Mapping (Grad CAM) and SHapley Additive exPlanations (SHAP), have been incorporated to generate interpretable visual explanations. The reviewed frameworks are evaluated on classification performance, robustness under heterogeneity, and stability of generated explanations. However, this review reveals significant gaps: the types of heterogeneity are addressed in isolation, XAI evaluation remains largely qualitative, and explanation stability under non-IID conditions lacks rigorous validation. These findings collectively highlight the need for federated frameworks that unify heterogeneity handling across all its forms simultaneously rather than addressing each in isolation, quantitative XAI assessment, and validation of explanation consistency across diverse hospital environments to enable trustworthy and interpretable clinical deployment.
Published by: Muhammad Auwal Yusuf
Author: Muhammad Auwal Yusuf
Paper ID: V12I3-1172
Paper Status: published
Published: May 26, 2026
Edge-Optimized Pre-Trained Deep Learning Models for Real-Time Detection of Red Palm Weevil and Date Palm Diseases: A Review
Date palm (Phoenix dactylifera L.) constitutes one of the most economically and culturally significant crops across arid and semi-arid regions, yet its productivity faces existential threats from the Red Palm Weevil (Rhynchophorus ferrugineus, RPW) and a spectrum of fungal and bacterial diseases. While deep learning has demonstrated remarkable classification accuracies exceeding 97% in controlled laboratory environments, the transition from academic prototypes to deployable, real-time agricultural solutions remains critically underdeveloped. This comprehensive review systematically examines recent studies, synthesizing the current landscape of deep learning applications for RPW detection and date palm disease classification. Our analysis reveals a persistent disconnect between architectural sophistication and practical deployability. Furthermore, the literature exhibits a pronounced fragmentation between pest detection and disease classification, with few studies addressing the integrated palm health ecosystem. This review identifies critical research dimensions where the current state-of-the-art falls short. By mapping these interconnected gaps across the evaluated literature, this review establishes a structured roadmap for developing lightweight, accurate, and interpretable AI systems that bridge the gap between theoretical accuracy and operational feasibility in precision agriculture.
Published by: Umar Faruk Ibrahim
Author: Umar Faruk Ibrahim
Paper ID: V12I3-1170
Paper Status: published
Published: May 26, 2026
Robust Deep Residual Networks with Pixel-Level Pre-Processing for Decentralized Traffic Sign Recognition
While traffic sign recognition systems play a vital role in road safety and autonomous driving, traditional architectures often suffer severe accuracy degradation under adverse environmental conditions such as low light, fog, and heavy shadows. Although federated deep learning and convolutional neural networks (CNNs) have successfully advanced decentralized edge intelligence, standard RGB image processing remains a critical bottleneck for vehicles encountering environmental noise. To address this, we propose a lightweight, decentralized ResNet-34 architecture designed for embedded applications, enhanced by a robust multi-space pixel-level pre-processing pipeline. By incorporating localized edge contrast enhancement and chromatic variance stabilization (utilizing HSV and Ohta spaces), the proposed system isolates critical luminance and structural features prior to decentralized feature extraction. The framework was trained and evaluated on the German Traffic Sign Recognition Benchmark (GTSRB) and the Belgian Traffic Sign Data Set (BTSD). The results demonstrate that coupling dynamic image pre-processing with federated residual learning yields a highly efficient, accurate, and environmentally resilient system suitable for real-time edge deployment.
Published by: Yenugurosireddygari Hemalatha, Sudhakar Bathala
Author: Yenugurosireddygari Hemalatha
Paper ID: V12I3-1194
Paper Status: published
Published: May 25, 2026
Transformer-Based Object Detection Architectures for Autonomous Driving Perception: A Comprehensive Review
Autonomous vehicle perception is one of the most important components in intelligent transportation systems, and the reliable trade-off between high fidelity of detection precision and computational efficiency in real time remains an open problem. Deep learning has proven to be very accurate in controlled settings, but bringing CNN-based solutions to deployment with high latency and substantial memory overhead is often a challenge to the end-to-end deployed Transformer solution. This thorough review provides a systematic analysis of recent developments in transformer-based detection architectures, consolidating 2024–2026 transformer- and CNN-based architectures for detection. It is a thorough review that systematically analyzes recent transformer-based detection architectures, summarizing the current transformer- and CNN-based detection architectures from 2024 to 2026. From our analysis, we can see that there is a clear lack of theoretical sophistication and the real-life edge-deployability of the hardware. In addition, there is a clear disconnection between the 2D camera-based detection approach and the 3D multimodal fusion approach in the literature. The critical research dimensions that are not well met by the current state-of-the-art are identified in this review, including small object detection in dense urban environments and robust inference under challenging weather conditions. This review provides a structured path forward by mapping these interrelated gaps and paving the way for the creation of lightweight, accurate and robust transformer detectors that can be deployed on their own in the field.
Published by: Muhammad Hamza
Author: Muhammad Hamza
Paper ID: V12I3-1176
Paper Status: published
Published: May 25, 2026
Do ESG Rating Divergences Predict Stock Underperformance?
While the ESG investing trend has shifted to the forefront, a rather worrying paradox is also becoming ever more evident, that of the dramatically divergent ESG evaluations rendered for the identical companies, consistently and by rating agencies across the industry. This paper investigates whether that difference of opinion, particularly when combined with unambiguous and highly optimistic environmental rhetoric in corporate filings, might be used as a quantifiable indicator of greenwashing, and if companies with that pattern of behaviour tend subsequently to underperform in equity markets. I employ a sample of 135 international firms from the S&P 500 and MSCI world indices, 2015-2023, providing a dataset of 1,080 firm-year observations. I calculated an aggregated ESG divergence index using pairwise disagreements from MSCI, Sustainalytics, and Bloomberg ESG ratings, and combined it with two text-based indicators from annual reports: a FinBERT sustainability sentiment index and a TF-IDF-based ESG keyword intensity measure for inclusion in my analysis. Ordinary Least Squares (OLS) regressions, two-way fixed effects panel models, and Fama-MacBeth cross-section estimation are used to carry out the empirical investigation. I find, throughout all specifications, that ESG rating divergence is negatively and significantly associated with risk-adjusted returns; each unit of added divergence relates to an annual excess return that is approximately 0.39–0.42% lower (p<0.01). Positive sustainability sentiment in disclosures correlates with better performance, whereas high keyword density unaccompanied by external rating agreement points in the opposite direction, consistent with rhetorical inflation rather than genuine ESG progress. A long-short portfolio sorted on divergence quintiles accumulates approximately 8.7 percentage points of excess return over the nine-year window. The results speak directly to the concerns of asset managers, index providers, and regulators engaged in the ongoing effort to bring rigour to sustainable finance.
Published by: Sheena Syed
Author: Sheena Syed
Paper ID: V12I3-1169
Paper Status: published
Published: May 22, 2026
Review of AI-Driven Methods to Enhance Search and Rescue Operations after Earthquakes
The devastating results of seismic events affect many people’s lives every time they occur. The main damage is caused to people being trapped under the rubble for long periods of time because rescue teams lack intelligent and prioritised plans. To address this problem, we will study the utilisation of various artificial intelligence (AI) technologies in accelerating and enhancing the efficiency of human spotting and rescuing processes after earthquakes. These applications include neural networks, AI-controlled robot systems, and multimodal data fusion, where thermal, visual, and acoustic data are integrated to detect survivors under building debris, in addition to other fusion ideas. This review will focus on studies published between 2023 and 2026. One main outcome of this work is the identification of key limitations in current AI solutions, such as real-time processing constraints, the noisy nature of disaster environments, and hardware deployment challenges in active disaster zones. In conclusion, this review aims to serve as a future research direction for optimising AI-driven methods to enhance detection accuracy, accelerate rescue timelines, and improve the overall survival outcomes in seismic emergencies.
Published by: Lin M.Saleh Aouto
Author: Lin M.Saleh Aouto
Paper ID: V12I3-1179
Paper Status: published
Published: May 22, 2026
