Deep learning for giant cell arteritis diagnosis on temporal artery biopsy.

OBJECTIVES

Giant Cell Arteritis (GCA) is a vasculitis affecting large and medium-caliber arteries, requiring early and accurate diagnosis to prevent serious complications. Temporal artery biopsy (TAB) is the gold standard for histopathological diagnosis, but its evaluation is challenging, time-consuming, and requires significant expertise. This study aimed to assess the accuracy of a deep learning model in diagnosing GCA from TAB images.

METHODS

We assembled a training cohort of 366 patients (137 GCA, 229 controls) and an external testing cohort of 58 patients (21 GCA, 37 controls). All whole-slide images (WSI) of hematoxylin-eosin-saffron-stained sections were digitized. Using CTransPath for feature representation, we trained a deep learning model with an {Updating}attention-based multiple-instance learning mechanism. The attention scores were analyzed to provide insights into the model's decision-making process.

RESULTS

The models achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.987 (±0.018) in 5-fold cross-validation on the training cohort. With a single misclassified case, they achieved a mean AUROC of 0.994 (±0.001) on the external testing cohort. The attention mechanism identified inflammatory infiltrates in the media layer as the primary features influencing predictions, demonstrating the model's interpretability.

CONCLUSION

Our deep learning framework achieved state-of-the-art performance in GCA diagnosis from TAB images. By accurately identifying critical features, it has the potential to assist pathologists in prescreening, reducing workload and improving diagnostic consistency.