Comprehensive Summary
Kondaveeti and Simhadri evaluated the reliability of deep learning models for detecting rice leaf diseases by combining traditional performance metrics with explainable artificial intelligence (XAI). Eight pre-trained convolutional neural networks (ResNet50, DenseNet201, InceptionV3, etc.) were tested on an augmented dataset of four rice leaf diseases. Models were assessed in three stages: classification accuracy, visualization of decision-making using LIME, and quantitative analysis of how well model-focused regions overlapped with expert-labeled diseased areas. ResNet50 performed best, with a 99.13% classification accuracy, strong feature alignment (IoU 0.432), and the lowest overfitting ratio (0.284). By contrast, InceptionV3 and EfficientNetB0, despite high accuracy, focused more on irrelevant features, raising concerns about reliability.
Outcomes and Implications
While the study centers on agriculture, its framework has broader biomedical relevance. Many medical AI systems face similar issues, they achieve high accuracy but sometimes rely on spurious features, undermining clinical trust. The proposed three-stage methodology—evaluating both prediction accuracy and interpretability—can guide development of safer, more transparent AI for healthcare diagnostics, such as radiology or pathology. By quantifying feature relevance and introducing an overfitting ratio, the approach could help clinicians and researchers identify models that not only classify correctly but also base decisions on meaningful biological patterns, a critical step for real-world deployment.