Rapid and automated detection of rice diseases is essential for minimizing crop losses. However, pixel-wise annotation of rice disease datasets is time-consuming and labor-intensive. This study proposed an iterative method to progressively improve the dataset’s quality by utilizing model-generated predictions. Initially, a model is trained on a limited-quality dataset and then used to relabel the training and validation sets. Predictions with low IoU compared to existing annotations but high confidence scores are flagged as potentially missing ground-truth instances. These are manually reviewed and corrected by human annotators before retraining the model on the updated dataset. Experimental results show consistent performance improvements across YOLOv8 variants, with YOLOv8s improving their F1-score from \(77.1\%\) to \(79.4\%\) and mAP \(_{50}\) from \(80.5\%\) to \(82.2\%\) after applying IPLC. Similarly, the Mask R-CNN models also benefit from IPLC, with the ResNet-50 backbone achieving an AP \(_{50}\) increased from \(70.74\%\) to \(76.34\%\) , and AP \(_{50:95}\) from \(44.93\%\) to \(49.36\%\) . These results demonstrated IPLC’s potential to enhance annotation quality and segmentation accuracy in deep learning models.