Banff score: from clinical practice to computer-aided tools

The BANFF classification is a comprehensive system that strives to establish uniformity in the lesions resulting from rejection sub sequent to renal transplantation. This process is particularly time-consuming and energy-intensive for pathologists, as each sample from the biopsies must be analyzed in detail. The aim of this project is to automate this process, identifying the five most relevant element namely glomeruli, blood vessels, tubules, nuclei, and interstitial tissue. In this work, we focus on the detection of tubules, vessels and nuclei using a new database composed by nineteen wsi images collected by the University Hospital Center (CHU) of Nice. From the original data base, we selected and labeled a total of fifty-six 1024 x 1024 patches, nine per patient. Using the selected data we were able to obtain a train set, a validation set and a test set sufficiently representative of the entire population under analysis. We propose two new pipelines for the segmentation of blood vessel and tubule lumens and nuclei. Both approaches were based on global thresholds and color deconvolution using PAS-stained images. The two algorithms include original pre- and post-processing strategies to improve the quality of the segmentation. To classify the structures, supervised machine learning-based methods were utilized, with three classifiers implemented: the K-Nearest Neighbors (KNN), the Random Forest (RF), and a Support Vector Machine (SVM). The novelty of our work lies in the method used to perform feature selection and feature extraction. We didn’t use conventional and automated techniques. To replicate the visual classification process carried out by a human we worked in in accordance with two expert pathologist and we conceive original methods for extract and selected the optimal set of features. In particular, the chosen feature were the size, width, height and eccentricity of the lumens. The difference between the convex hull evaluated on the lumen and the mask of the lumen itself. The number of nuclei connected to each lumen and the mean and the standard deviation of the size and eccentricity of these nuclei. We obtain an accuracy of 91.03% and a balanced accuracy of 91.08% on the test set using the SVM classifier. With regards to tubules, the precision was 92.80% and the recall was 89.58%. For vessels, the precision was 89.28% and the recall was 92.59%. Furthermore, to test the robustness of the method we applied the best performing model to set of six images selected by a skilled pathologist. These images are intended to represent the worst possible case. To make the model free from color dependency we also applied the best model on some TRI-stained images. In both cases we get acceptable performance. Although embryonic, this thesis work lays the foundations for the automation of the BANFF classification providing an essential tool for increasing the number of diagnoses.