P0119ARTIFICIAL INTELLIGENCE IN RENAL PATHOLOGY: IBM WATSON FOR THE IDENTIFICATION OF GLOMERULOSCLEROSIS

Abstract

Abstract Background and Aims Artificial Intelligence (AI) is the branch of computer technology aimed at creating hardware and software systems solving problems similarly to human intelligence, whereas specifically machine learning (ML) is the “field of study that gives computers the ability to learn without being explicitly programmed”. Our study was carried out taking advantage of the Watson Visual Recognition System by IBM, an advanced AI tool based on ML able to classify complex visual content. The aim of the study was to train and test the ability of the system in recognizing sclerotic and non-sclerotic glomeruli. Method A dataset of 26 renal biopsies performed at the Nephrology unit of the Department of Emergency and Organ Transplantation (DETO), University of Bari, Italy, was used for the analysis. All biopsies were stained with Periodic Acid-Schiff (PAS) staining. Each bioptic section was acquired using Aperio ScanScope; glomeruli were identified and sclerotic glomeruli were marked in yellow, non-sclerotic ones in green. Annotations were validated by two renal pathologists. The final dataset consisted of 2772 glomeruli: 428 with sclerosis, 2344 with no sclerosis (ratio 1/5.5). The dataset was divided in three parts: training set (about 70% of the entire dataset), validation set (about 10%) and test set (about 20%). Watson Visual Recognition Service is customizable and the system can be trained on recognizing any visual content. Classifiers are created and trained uploading both positive (in our study, sclerotic glomeruli) and negative (non-sclerotic glomeruli) classes. The IBM Watson learning algorithm is not open, therefore in order to improve the performance of the system, it is necessary to train it with different models and choose the best one. Results We created all the possible models derived from the arrangement of the following 4 variables: color of the image (PAS staining or grey scales), dimension (original or resized), number of images in order to balance the positive and negative classes, binary (one class containing sclerotic glomeruli) or multi-class (two classes: sclerotic and non-sclerotic glomeruli). Every test had a cut-off of 0.5: if >0.5, the system considered the glomerulus belonging to the tested class. After validating all the models with all the variables considered, the best performing model was the following: grey scaled, resized and multi-class. This model was then tested with a different number of input images (300, 600, 900, 1200, and 1600). The models with the most numerous dataset have been created using data augmentation, a technique that virtually increases the number of available samples. Results show that the use of larger input datasets does not yield a better linear performance. In fact, models 900 and 1200 had worst performances than other models, the best performances of the system were reached with model 1600, both in recognizing sclerotic (positive class) and non-sclerotic (negative class) glomeruli (Table 1, Figure 1 and 2). Conclusion In our study, renal biopsy images were analyzed and classified using the IBM Watson Visual Recognition tool which was able to distinguish automatically and with very high accuracy between sclerotic and non-sclerotic glomeruli. This study focuses only on the glomerular compartment and the next step will be the recognition of intermediate lesions and other portion of renal tissue. Our results prove the potential of AI and ML techniques in supporting the activity of renal pathologists.