*Purpose: Arterial intimal fibrosis (AIF) is the excessive fibrotic tissue formation in the intima of arteries indicative of poor graft function after kidney transplant. Visual assessment of AIF by the Banff criteria is not consistent among pathologists due to inter- and intra-observer variation. Seeking objective and accurate quantitative analysis, this study built a deep-learning (DL) model that automatically detected arteries and AIF in Periodic acid-Schiff (PAS)-stained whole slide images (WSI) of kidney transplant biopsies.

*Methods: 120 WSIs were annotated with arteries and AIF and split as training (n=78 slides) and testing (n=42 slides) sets. The annotated sections were pre-processed with data augmentation to 880 fix-sized images. Our DL model was built on the training set with 5-fold cross-validation based on Mask R-CNN structure. Established model was applied to testing set and detection accuracies were determined by true positive rate (TPR) and positive predictive value (PPV) on instance- and pixel-level. The slide-wide AIF level was extracted and correlated with Banff cv score.

*Results: The model generated predicted masks for arteries and AIF of all test images, and resulting masks were stored in run-length encoded (RLE) format for efficient storage. On the instance-level, the TPR and PPV of the model’s artery detection were 0.82 and 0.85; the TPR and PPV of AIF were 0.73 and 0.75. On the pixel-level, the TPR and PPV of artery detection and segmentation were 0.84 and 0.83; and the TPR and PPV of AIF were 0.71 and 0.72. Spearman’s correlation coefficient between %AIF in arteries of all WSI and their associated Banff cv scores was 0.501 with a significant p-value of 1.4E-8, showing a general positive correlation.

*Conclusions: The DL model more accurately detects and quantifies arteries and AIF in pathological cases, meanwhile allowing objectivity and reducing processing time compared to the subjective and time-consuming assessment by the Banff criteria. The model also better detected complete arteries with closed circular shapes than incomplete or cut arteries in WSI, showing room for model improvement to identify AIF and arteries with uncommon shapes. In clinical settings, our automated model can provide efficient and accurate diagnostic suggestions to pathologists that streamlines kidney allograft pathology classification.

« Back to 2022 American Transplant Congress