Accurate and timely detection of crop diseases is essential for global food security and sustainable agriculture. This review provides a comprehensive analysis of recent advancements in leaf image-based crop disease detection using machine learning, deep learning (DL), convolutional neural networks, vision transformers, and emerging state-space models. Following PRISMA guidelines, the review systematically examines peer-reviewed studies (2021–2025) and evaluates datasets, model architectures, validation protocols, and deployment constraints. Unlike previous surveys, this review uniquely integrates explainable artificial intelligence (XAI) with emphasis on physiology-based interpretability and real-world field usability. Key findings reveal that while DL models achieve > 95% accuracy on controlled datasets like PlantVillage, performance degrades significantly under field conditions due to dataset bias, domain shifts, and multi-disease complexity. Critical research gaps are identified, including insufficient annotation standards, limited cross-domain validation, and computational constraints for edge deployment. Emerging directions include agriculture-specific foundation models, domain-generalizable vision systems, lightweight mobile strategies, and next-generation XAI frameworks aligned with plant physiology. By bridging technical advancements with agronomic interpretability, this work provides researchers and practitioners with a roadmap for developing trustworthy, field-deployable AI systems for sustainable crop health management. Graphical abstract The alternative text for this image may have been generated using AI.

Background: The proposed Architecture will provide the processing and analysis essential to accurate and reliable detection of brain tumors from MRI, for timely diagnosis and evidence-based decisions. Medical imaging now routinely enters clinical assessment; the thrust is shifting toward attaining high performance within open and governed systems to enable deployment in real-world healthcare applications. Methods: This paper proposes a two-stage deep learning framework: first, DeepLabV3 for segmentation to demarcate candidate tumor regions, and then CNN to classify whether a tumor exists. The different components employ pre-trained models through transfer learning and fine-tuning. The DeepLabV3 and CNN architectures are used, together with metric computation modules. This approach will be tested on BraTS MRI data. For efficient model training, optimizers such as SGD, RMSprop, and Adam can be employed. Results: The classification performance could be achieved with a high value of 99.31% using an ADAM optimizer in the proposed architecture. Besides, both the precision and recall are very high, indicating good generalization and stable performance. Moreover, segmenting before classification provides more reliable detection compared to using a single-stage model. Discussion: These results indicate that feature learning guided by segmentation enhances tumor detection with a binary classifier, while remaining interpretable and robust. This makes the framework much more transparent and easy to audit, suitable for use in cloud-enabled, secure, and IoT-enabled clinical environments. Originality: It therefore proposes a two-layer deep learning architecture that effectively incorporates precise tumor localization into explicit binary tumor detection. Beyond this, the work focuses on practical clinical applicability, robust data governance, and deployment-ready systems rather than diagnosing subtypes of tumors.