- Abstract: Pathologic scoring on renal allograft specimen mainly depends on pathologists’ visual scoring, which would be time consuming and sensitive to inter or intra-observer variations. In this study, our aims are two folds; one is to propose a fully-automated system to find feasible regions of interest (ROIs) and count their number of C4d positive and negative in PTC on each feasible ROI by convolutional neural net (CNN) method in the giga-pixel immunostaining pathologic slide images. The other is to validate whether AI-assisted labeled data by a detection network is feasible. Our results showed that the performance in terms of area under curve (AUC) for classification model to find feasible ROIs is 0.9601 and the sensitivities to detect positive and negative in PTC is 0.9413 and 0.8523 at mean of false positive (FP)s 1.5 and 4 per ROI, respectively. In addition, we proved that AI-assisted labeled data is feasible by showing that the sensitivities were increased to 0.9522 and 0.8864 at the same mean FPs, respectively.
- Author affiliation: Asan Medical Center & University of Ulsan College of Medicine
- Keywords: convolutional neural net (CNN), deep learning, immunostaining, peri-tubular capillary (PTC), renal allograft rejection