Abstract
Artificial intelligence (AI) has transformed medical imaging, notably in radiology and endoscopy. Semantic segmentation, a pixel-level technique crucial for delineating pathological features, has become pivotal in digital pathology. Pathology segmentation AI models are often trained using annotations generated by pathologists. Despite the meticulous care typically exercised, pathologist-generated annotations often contain label noise whose types and effects on model training remain underexplored. This study combined a survey of public datasets with the synthesis of artificial label noise to evaluate its effects on pathology segmentation models. Using publicly available datasets and a breast cancer semantic segmentation dataset, modules were developed to simulate four types of artificial label noise at varying intensity levels. These datasets were used to train deep learning models and their performance was evaluated. The results indicated that models were highly susceptible to overfitting label noise, particularly boundary-dependent noise, such as dilation and shrinkage. Discrepancies were identified between apparent performance scores obtained under real-world conditions and true performance scores derived using clean test data. This overestimation risk was most pronounced for datasets containing boundary-altering noise. Furthermore, random noise combinations further degraded generalization. This study underscores the critical importance of addressing label noise in pathology datasets. It is proposed that future efforts focus on developing standardized methods for quantifying and mitigating label noise, along with creating robust benchmarks using noise-inclusive datasets. Enhancing annotation quality and addressing label noise can improve the reliability and generalizability of AI in pathology, facilitating broader clinical adoption.