Detection and Segmentation of Kidneys from Magnetic Resonance Images in Patients with Autosomal Dominant Polycystic Kidney Disease

The segmentation of kidneys from Magnetic Resonance (MR) images of subjects affected by Autosomal Dominant Polycystic Kidney Disease (ADPKD) is a task of fundamental importance as it allows a non-invasive assessment and monitoring of the disease over time. In this work, a fully automated procedure based on Convolutional Neural Networks (CNNs) is proposed for detecting images containing the kidney and subsequently segment them classifying each pixel. Specifically, a mono-objective genetic algorithm was designed for optimising the CNN architecture, modelling the number of encoders, the structure of each encoder and the final fully-connected layers. The input dataset for the classification task included 526 MR images: 366 containing kidney and 160 did not include any pixel of the kidney. All the images containing the kidney were split into left and right side and used as input dataset for the segmentation procedure. For both the tasks, the training set, validation set and test set were generated according to the 60-20-20 percentages. Accuracy higher than 95% for the classification task and 90% for the segmentation algorithm show the reliability of the proposed approach, also improving the results of previous works.