Abstract
Purpose :
This pilot study aims to develop a deep residual convolutional network (ResNet) model for image tile/patch classification combined with a whole-slide inferencing mechanism for determining predominant and minor giant cell arteritis (GCA) histologic patterns.
Methods :
Hematoxylin and eosin (H&E) stained, formalin-fixed, paraffin-embedded tissue specimens from 1946 patients who underwent temporal artery biopsies (TAB) from January 1, 2000, to December 31, 2019, at the West Virginia University Hospitals were de-identified and digitized using a whole-slide digital scanner at 20x magnification. Digital scans of whole slides were chronologically partitioned into two sets, the first 80% samples for model training and the remainder as an independent test set to validate the algorithm. Based on identified definitions/characterizations, several image tiles or patches from each scan were identified as regions of interest (ROIs) for GCA detection and manually labeled for training the deep learning model. These tiles were resized into square patches. The developed ResNet classification model takes square patches or tiles as inputs and predicts the probability of GCA presence.
Results :
The dataset contained 693 positive and 1253 negative training tiles. We performed training trials after expanding the same ROI image set by augmentation. We applied six transformations to each training image: Rotate 90, Vertical Flip, Horizontal Flip, Rotate Randomly between +10 and -10 degrees, Color Jitter, and Histogram Equalization and obtained 100% accuracy on the running train accuracy as well as the validation set. The algorithm was tested on unseen data and performed consistently with an AUC of 0.99 at the whole tissue section level. Model prediction and pattern detection were qualitatively validated using a class-activation map visualization method called GradCAM.
Conclusions :
Deep neural network learning methods help automate the detection of GCA in TAB digital pathology slides. The findings generated by the model are comparable in performance to an experienced pathologist.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.