Automated detection of giant cell arteritis from temporal artery biopsy specimens using deep learning approaches

This pilot study aims to develop a deep learning model for classifying temporal artery biopsy (TAB) histological sections to detect histologic patterns indicative of giant cell arteritis (GCA). Formalin-fixed, paraffin-embedded, hematoxylin and eosin (H&E) -stained, tissue specimens from 472 patients who underwent TAB between January 1, 2000, and December 31, 2019, were digitized at 20x magnification. Individual artery regions were identified, extracted, and resized into individual image patches/tiles, referred to as regions of interest (ROIs), for GCA detection. A ResNet model was trained using these ROIs after data augmentation techniques. Performance metrics such as accuracy and area under the receiver operating characteristic curve (AUC) were used to evaluate the model. The training set included 336 slides (100 positive, 236 negative), and the test set comprised 136 slides (40 positive, 96 negative). The ResNet model achieved an accuracy of 96.32% with an AUC of 0.99 on the validation set, and 92.32% accuracy with an AUC of 0.93 on a held-out test set. Model predictions were validated using GradCAM visualizations which qualitatively confirmed the model’s performance. This study demonstrates the effectiveness of deep neural network methods in automating the detection of GCA from TAB, and this approach holds promise for speeding up diagnosis and improving test sensitivity.