This M.Sc. thesis studies deep-learning models for kidney failure prediction from renal biopsy imaging.

Abstract

The expression renal failure indicates inability of the kidneys to perform excretory function leading to retention of nitrogenous waste products from the blood. Acute and chronic renal failure are the two genre of kidney failure. When a patient needs renal replacement therapy, the condition is known as end-stage renal disease (ESRD) and it often needs dialysis, that is a process used to remove refuses and other substances from the blood. Many patients are asymptomatic and are casually found to have an important serum creatinine concentration, extraordinary urine studies (such as proteinuria or microscopic hematuria), or abnormal radiologic imaging of the kidneys. Deep learning uses convolutional neural networks (CNNs), artificial correspective of neural network in human brain, enabling computer learning complex predictive tasks without directly programming, the algorithm is exposed to desired input-output behavior, from which it extracts features that are useful to learn how to predict the future state of the kidney. In this work we deepen the effectiveness of state-of-art deep learning models applied on images that have very high resolutions and usually lack localized annotations on nephropathy patients. The purpose of this current study is to investigate and to enhance the deep learning prognostic score (DLPS) for nephropathy combining the results obtained using whole slide images (WSI) with the immunofluorescent counterpart to predict kidneys’ failure on nephropathy patients.

Details

Italian title: Previsione dell’Insufficienza Renale con Reti Neurali Combinando Immagini WSI e in Immunofluorescenza.

Supervisors listed by MoreThesis: Costantino Grana and Federico Bolelli.

Keywords: computer vision, deep learning, medical imaging, neural networks, whole-slide images.

---

← Back to projects