AB034. Weakly supervised artificial intelligence-based subtyping of thymic epithelial tumors using H&E whole slide images

Abstract

Background: Thymic epithelial tumors (TETs) are rare neoplasms classified by the World Health Organization (WHO) into thymomas (A, AB, B1, B2, B3) and thymic carcinomas (TCs). Due to overlapping histopathological features, accurate differentiation is challenging, leading to diagnostic variability among pathologists. This study aimed to develop a diagnostic tool to support precise subtyping, reduce ambiguity, improve treatment strategies, and ultimately enhance patient outcomes. Methods: Our study utilized an in-house Erasmus Medical Center (EMC) dataset of 669 patients diagnosed by eight pathologists and an external Lyon dataset of 97 patients. After excluding cases with less than 70% agreement within the pathologist panel and irrelevant thymoma types (n=510), 159 EMC cases remained. Tumor areas were annotated, and 512×512-pixel tiles were taken at 10x magnification using QuPath. To reduce staining differences, Vahadane’s stain normalization method was used (see Figure 1). The dataset was divided at the patient level, resulting in 76 EMC cases and 12 Lyon cases for training, while 83 EMC cases and 85 Lyon cases were used for testing. After preprocessing, two separate AI models were developed: the first was designed to generate additional labels, while the second, built on the VGG16 architecture, was used to classify the subtypes A, AB, B1, B2, B3, and TC. To ensure the reliability of the results, stratified 3-fold cross-validation was employed. Results: The model achieved an area under the curve (AUC) of 1 and a balanced accuracy (BAcc) of 0.97±0.02 on the validation set. On the 70–100% consensus test set, it recorded an AUC of 0.91±0.01 and BAcc of 0.78±0.03, while the 100% consensus test set yielded an AUC of 0.98±0.01 and BAcc of 0.89±0.01. On the external dataset, it reached an AUC of 0.93±0.01 and BAcc of 0.76±0.04. Grad-CAM visualizations showed subtype-specific patterns, although distinguishing group B remained challenging. Conclusions: The model demonstrated robust accuracy in distinguishing between TET subtypes in both external and internal dataset, thereby providing substantial support to pathologists and facilitating informed treatment decisions. Nonetheless, similar to the difficulties faced by pathologists, the model’s efficacy was compromised in ambiguous cases characterized by mixed histological features. In future work, we will evaluate its performance on additional external.

Ähnliche Arbeiten