| 1. |
- Wetzer, Elisabeth, et al.
(author)
-
Towards automated multiscale Glomeruli detection and analysis in TEM by fusion of CNN and LBP maps
- 2019
-
In: 3rd NEUBIAS Conference. - Luxembourg.
-
Conference paper (other academic/artistic)abstract
- Glomeruli are special structures in kidneys which filter the plasma volume from metabolic waste.Podocytes are cells that wrap around the capillaries of the Glomerulus. They take an active role in the renal filtration by preventing plasma proteins from entering the urinary ultrafiltrate through slits between so called foot processes. A number of diseases, such as minimal change disease, systemic lupus and diabetic nephropathy, can affect the glomerulus and have serious impact on the kidneys and their function.When the resolution of optical microscopy is insufficient for a diagnosis, it is necessary to thoroughly examine the morphology of the podocytes in transmission electron microscopy (TEM). This includes measuring the size and shape of the foot processes, the thickness and overall morphology of the Glomerulus Base Membrane (GBM), and the number of slits along the GBM.The high resolution of TEM images produces large amounts of data and requires long acquisition time, which makes automated imaging and Glomerulus detection a desired option. We present a multi-step and multi-scale approach to automatically detect Glomeruli and subsequently foot processes by using convolutional neural networks (CNN). Previously, texture information in the form of local binary patterns (LBPs) has shown great success in Glomerulus detection in different modalities other than TEM. This motivates our approach to explore different methods to incorporate LBPs in CNN training to enhance the performance over training exclusively on intensity images. We use a modified approximation of the Earth mover’s distance to define dissimilarities between the initially unordered binary codes resulting from pixel-wise LBP computations.Multidimensional scaling based on those dissimilarities can be applied to compute LBP maps which are suitable as CNN input. We explore the effect of different radii and dimensions for the LBP maps generation, as well as the impact of early, mid and late fusion of intensity and texture information input. We compare the performance of ResNet50 and VGG16-like architectures. Furthermore we provide comparison of transfer learning of networks pretrained on ImageNet, as well as on a publicly available SEM database, a network architecture in which convolutional layers are replaced by local binary convolutional layers, as well as ‘classic’ methods such as SVM or 1-NN classification based on LBP histograms.We show that for Glomerulus detection, where texture is a main discriminative feature, CNN training on the texture based input provides complementary information not learned by the network on the intensity images and mid and late fusion can boost performance. In foot process detection, in which the scale shifts the focus from texture to morphology, the performance also benefits by the handcrafted texture features, though to a lesser extent than for the larger scale Glomerulus detection.
|
|
