AI-Enhanced Subtyping of Thymic Tumors: Attention-based MIL with Pathology-Specific Feature Extraction

Abstract

ABSTRACTPurposeThe precise classification of thymic tumors using whole slide images (WSIs) is essential for accurate diagnosis and treatment. While traditional Convolutional Neural Networks (CNNs) are commonly used for this purpose, emerging models tailored to pathology, such as Phikon and HistoEncoder, present promising alternatives as feature extractors. Additionally, the limited availability of annotated WSIs has driven the development of weakly-supervised classifiers like multiple-instance learning (MIL) models. In this study, we evaluate nine different combinations of extractor-classifier pairs for thymic tumor subtyping, including a novel, self-developed attention-based MIL classifier, AttenMIL.MethodsThe process began with curating a dataset of thymic tumor Whole Slide Images (WSIs) from the TCGA platform. Using the Yottixel method, patches were derived from these WSIs, and features were extracted from the patches using three different pathology-specific models: Phikon, HistoEncoder, and a pathology-fine-tuned ResNet50. The extracted features were then organized into small bags of instances through a chunking technique. Subsequently, three MIL classifiers AttenMIL, TransMIL, and Chowder were trained. Finally, the efficacy and generalizability of nine different combinations of extractor-classifier pairs were evaluated on unseen test images. Confusion matrices for each pair were utilized to provide insights into misclassification patterns and potential error sources.ResultsThe Phikon feature extractor consistently delivered the highest classification accuracies, particularly when paired with the AttenMIL and Chowder classifiers, achieving up to 99% accuracy. This combination significantly outperformed other feature extractor-classifier pairs. Confusion matrices revealed that the AB and B3 subtypes were the most commonly confused classes across the different models.ConclusionsThe study demonstrates the potential of domain-specific feature extractors like Phikon, when coupled with robust MIL classifiers such as the novel AttenMIL and Chowder, in enhancing the accuracy and reliability of thymic tumor classification. The chunking-based augmentation method proved effective for thymic tumors, which are relatively homogeneous, but its applicability to heterogeneous tumors remains to be explored. Future research should address class imbalances and improve generalizability to different datasets.Statements and DeclarationsCompeting InterestsThe author has declared no competing interest.No funding was received for conducting this study.Declaration of generative AI and AI-assisted technologies in the writing processWhile preparing this work, the author used ChatGPT to improve language and readability. After using this tool/service, the author reviewed and edited the content as needed and took full responsibility for the publication’s content.