AB037. AI-powered differentiation of thymomas (A and B3) and thymic carcinoma using histopathological analysis of H&E whole slide images

Background: Thymic epithelial tumors (TETs), classified by the World Health Organization (WHO) into thymomas (A, AB, B1, B2, B3) and thymic carcinoma (TC), pose diagnostic challenges due to their morphological similarities. Distinguishing between thymomas (particularly A and B3) and TC is critical for tailored treatment and prognostic assessment. Despite B3’s classification as a well-differentiated TC, its resemblance to certain thymoma subtypes complicates accurate diagnosis. Our objective is to develop an artificial intelligence (AI) model that is adept at differentiating thymomas (A and B3) from TC, leveraging computational methods to enhance diagnostic accuracy and design appropriate treatment plans.

Methods: A binary convolutional neural network (CNN) classifier was developed to distinguish between thymoma (A-B3) and TC. The dataset comprised 31 patients diagnosed with TET by eight pathologists. The cases included 13 type A, 5 type B3 thymoma cases, and 13 TC cases. The hematoxylin and eosin (H&E) whole slide images (WSIs) were annotated using QuPath, and 512×512-pixel patches were extracted at 10× magnification. The images underwent stain normalization using the Vahadane method. The model underwent training and validation using stratified k-fold cross-validation (K=3), involving 21 patients for train-validation, and 10 patients for testing.

Results: Our classification model showed robust performance with a patient-level area under the curve (AUC) of 0.86±0.1 on validation and 0.86±0.04 on the test set. Notably, it achieved perfect AUC of 1 in validation and 0.92 in testing, with balanced accuracies of 1 and 0.91, respectively, identifying the best model in fold 3. However, it struggled to predict thymoma (A-B3) samples with lymphocyte densities similar to or higher than those in TC samples. GradCAM was used to visualize influential features affecting these predictions.

Conclusions: The proposed model effectively distinguishes thymoma (A-B3) from TC, demonstrating high generalizability in k-fold outcomes at the patient level. Accurate subtyping is crucial as TCs typically require chemotherapy, while thymomas are managed with surgical resection and potentially radiotherapy, but never chemotherapy. These findings underscore the importance of precise diagnoses to enhance treatment precision and prevent unnecessary chemotherapy in TET patients, highlighting the significant role of accurate classification in patient management.

Keywords: Thymic epithelial tumor (TET); artificial intelligence (AI); digital pathology