Accurate and timely diagnosis of rice leaf diseases is crucial for smart agriculture leveraging vision sensors. However, existing lightweight convolutional neural networks (CNNs) often struggle in complex field environments, where small lesions, cluttered backgrounds, and varying illumination complicate recognition. This paper presents I-GhostNetV3, an incrementally improved GhostNetV3-based network for RGB rice leaf disease recognition. I-GhostNetV3 introduces two modular enhancements with controlled overhead: (1) Adaptive Parallel Attention (APA), which integrates edge-guided spatial and channel cues and is selectively inserted to enhance lesion-related representations (at the cost of additional computation), and (2) Fusion Coordinate-Channel Attention (FCCA), a near-neutral SE replacement that enables efficient spatial–channel feature fusion to suppress background interference. Experiments on the Rice Leaf Bacterial and Fungal Disease (RLBF) dataset show that I-GhostNetV3 achieves 90.02% Top-1 accuracy with 1.831 million parameters and 248.694 million FLOPs, outperforming MobileNetV2 and EfficientNet-B0 under our experimental setup while remaining compact relative to the original GhostNetV3. In addition, evaluation on PlantVillage-Corn serves as a supplementary transfer sanity check; further validation on independent real-field target domains and on-device profiling will be explored in future work. These results indicate that I-GhostNetV3 is a promising efficient backbone for future edge deployment in precision agriculture.

Medical-image annotation is costly and limits the use of fully supervised learning. This study proposes the Self-Supervised Stacked Masked Denoising Autoencoder (S²MDAE) for four-class brain MRI classification. During pre-training, three convolutional encoder–decoder blocks perform layer-wise masked reconstruction under a 75% mask ratio and Gaussian corruption; the pretrained encoder stack is then fine-tuned for classification. On the primary dataset, S²MDAE achieved 87.207% Accuracy, while ablation studies supported the contributions of pre-training, masking, noise injection, and layer-wise reconstruction. Representation analysis further showed increased class separability in deeper blocks after fine-tuning. Without retraining, external evaluation on a duplicate-screened independent dataset achieved 67.254% Accuracy and 66.009% Macro-F1, indicating partial transfer under domain shift. These findings support the task-specific value of the proposed framework while showing that broader clinical generalization requires further validation.