Pethmunee Tanatorn, Kansomkeat Supaporn, Bhurayanontachai Patama, Intajag Sathit
Division of Computational Science, Faculty of Science, Prince of Songkla University, Songkhla 90110, Thailand.
Department of Ophthalmology, Faculty of Medicine, Prince of Songkla University, Songkhla 90110, Thailand.
Diagnostics (Basel). 2025 Jul 8;15(14):1735. doi: 10.3390/diagnostics15141735.
In automatic medical image analysis, particularly for diabetic retinopathy, the accuracy of labeled data is crucial, as label noise can significantly complicate the analysis and lead to diagnostic errors. To tackle the issue of label noise in retinal image segmentation, an innovative label correction framework is introduced that combines Confident Learning (CL) with a human-in-the-loop re-annotation process to meticulously detect and rectify pixel-level labeling inaccuracies. Two CL-oriented strategies are assessed: Confident Joint Analysis (CJA) employing DeeplabV3+ with a ResNet-50 architecture, and Prune by Noise Rate (PBNR) utilizing ResNet-18. These methodologies are implemented on four publicly available retinal image datasets: HRF, STARE, DRIVE, and CHASE_DB1. After the models have been trained on the original labeled datasets, label noise is quantified, and amendments are executed on suspected misclassified pixels prior to the assessment of model performance. The reduction in label noise yielded consistent advancements in accuracy, Intersection over Union (IoU), and weighted IoU across all the datasets. The segmentation of tiny structures, such as the fovea, demonstrated a significant enhancement following refinement. The Mean Boundary F1 Score (MeanBFScore) remained invariant, signifying the maintenance of boundary integrity. CJA and PBNR demonstrated strengths under different conditions, producing variations in performance that were dependent on the noise level and dataset characteristics. CL-based label correction techniques, when amalgamated with human refinement, could significantly enhance the segmentation accuracy and evaluation robustness for Accuracy, IoU, and MeanBFScore, achieving values of 0.9156, 0.8037, and 0.9856, respectively, with regard to the original ground truth, reflecting increases of 4.05%, 9.95%, and 1.28% respectively. This methodology represents a feasible and scalable solution to the challenge of label noise in medical image analysis, holding particular significance for real-world clinical applications.
在自动医学图像分析中，尤其是对于糖尿病视网膜病变，标记数据的准确性至关重要，因为标记噪声会使分析显著复杂化并导致诊断错误。为了解决视网膜图像分割中的标记噪声问题，引入了一种创新的标记校正框架，该框架将置信学习（CL）与人工参与的重新标注过程相结合，以精心检测和纠正像素级标记不准确之处。评估了两种面向CL的策略：采用带有ResNet-50架构的DeeplabV3+的置信联合分析（CJA），以及使用ResNet-18的按噪声率修剪（PBNR）。这些方法在四个公开可用的视网膜图像数据集上实现：HRF、STARE、DRIVE和CHASE_DB1。在模型在原始标记数据集上训练后，对标记噪声进行量化，并在评估模型性能之前对疑似误分类像素进行修正。标记噪声的减少在所有数据集的准确性、交并比（IoU）和加权IoU方面带来了一致的提升。微小结构（如中央凹）的分割在细化后显示出显著增强。平均边界F1分数（MeanBFScore）保持不变，表明边界完整性得以维持。CJA和PBNR在不同条件下表现出优势，性能变化取决于噪声水平和数据集特征。基于CL的标记校正技术与人工细化相结合时，可以显著提高分割准确性以及准确性、IoU和MeanBFScore的评估稳健性，相对于原始地面真值，分别达到0.9156、0.8037和0.9856的值，分别提高了4.05%、9.95%和1.28%。这种方法代表了一种可行且可扩展的解决方案，用于应对医学图像分析中标记噪声的挑战，对实际临床应用具有特别重要的意义。
