DEEP LEARNING TECHNIQUES FOR KIDNEY DISEASE DETECTION

Abstract

Around 10% of the world's population gets affected by Chronic Kidney Diseases or CKDs at some point in their lives and millions die each year due to not having access to affordable treatments and clinical facilities. During the last two decades, CKDs and short-term Acute Kidney Injuries (AKIs) have been steadily increasing in low- and middle-income countries due to obesity, diabetes, and other diseases. Based on recent studies, CKDs are directly correlated to kidney cancers and the diagnosis of kidney cancers through laboratory tests is time-consuming, complicated, unreliable, and costly. Kidney cancer is, in fact, one of the most common malignant CKDs worldwide. Accurate diagnosis is a critical step in the management of patients with kidney cancer and is influenced by multiple factors including tumor size or volume, cancer type or stage, etc. For malignant tumors, partial or radical surgery of the kidney might be required, but for many clinicians, the basis for making this decision is often unclear. Partial nephrectomy could result in avoidable patient death if kidney removal was needed, whereas total nephrectomy in less severe cases could resign patients to life-long dialysis or need for future transplantation without sufficient cause. Using machine learning to consider clinical data alongside computed tomography images could potentially help resolve some of these issues, by enabling a more robust classification of kidney cancers and selection of optimal surgical approaches. In this study, I used the publicly available KiTS dataset of contrast-enhanced CT images and corresponding patient metadata to differentiate four major classes of renal cell carcinoma: clear cell (ccRCC), chromophobe (chRCC), papillary (pRCC), and oncocytoma (ONC). I rationalized these data to overcome the high field of view (FoV), extract tumor regions of interest (ROIs), classify patients using deep machine learning models, and extract / post-process CT image features for combination with clinical data. Regardless of marked data imbalance, the proposed combined approach with the help of the proposed novel DenseAUXNet201 classifier achieved a high level of performance (85.66% accuracy, 84.18% precision, 85.66% recall, and 84.92% f1-score). When selecting surgical procedures for malignant tumors (RCC), the proposed method proved even more reliable (90.63% accuracy, 90.83% precision, 90.61% recall, and 90.50% f1-score). Using feature ranking, it was confirmed that tumor volume and cancer stage are the most relevant clinical features for predicting surgical procedures. Once fully mature, the approach I propose could be used to assist surgeons in performing nephrectomies by guiding the choices of optimal procedures in individual patients with kidney cancer. Adults are not the only ones who suffer from kidney complications, even fetuses might suffer from kidney problems such as hydronephrosis. Antenatal or prenatal hydronephrosis (AHN) is a common kidney complication in unborn children. While AHN is generally benign and resolves over time, in severe cases this condition can inflict serious kidney damage or even organ failure due to excessive waste accumulation. Regardless of severity, AHN must be clinically monitored for resolution, with treatment plans and medications being revised according to updated prognoses. Kidney ultrasound (US) images are one of the most common methods of monitoring AHN, but grading of this condition is highly subjective and clinicians may select inappropriate therapies or surgical interventions as a result. New approaches are required to differentiate subjects who can be managed without surgical intervention from those who require life-saving operations. An end-to-end deep machine learning framework was developed to sequentially detect ultrasound regions of interest (ROIs), segment kidneys from US images, and classify AHN severity. I propose the novel Kidney Ultrasound Segmentation Network (KUSNet) for kidney segmentation from ultrasound images, and the Prenatal Hydronephrosis Classification Network (PHCNet) for hydronephrosis severity stratification according to Society of Fetal Urology standards. At each stage, the performance of the proposed models was assessed both quantitatively and qualitatively against state-of-the-art networks in the respective fields. PHCNet achieved 87.7% accuracy, 88.1% precision, 87.7% recall, 78.0% specificity, and 87.7% f1-score when performing multiclass stratification of AHN severity. Artificial intelligence-based tools can reliably classify AHN severity to reduce inter- and intra-observer bias, thereby aiding clinicians in the rapid selection of appropriate treatments and surgeries.