Multi-scale attention network for segmentation of electron dense deposits in glomerular microscopic images

Electron dense deposit on the epithelial side of the glomerular capillary basement membrane is one of the pathological changes of membranous nephropathy. Automatic segmentation of deposits can relieve clinicians from the tedious and manual effort of identifying and localizing region of interest (ROI) in medical images and also assist to diagnose membranous nephropathy. Electron dense deposits are characterized by different sizes, irregular shapes, and low contrast to surrounding tissue structures in glomerular electron microscopy images. Considering the characteristics of dense deposits, we propose a multi-scale attention network for automatic segmentation of electron dense deposits of glomeruli in electron microscope images. Our method is built on the fully convolutional network but also takes advantages of the multi-scale skip connections and attention mechanism. Specifically, the multi-scale skip connection combines feature maps of different scales, makes the segmentation field larger, and integrates the shallow features of the image and high-level semantic information, which is more conducive to distinguishing dense deposits. At the same time, attention mechanism can focus on salient structures that normally produces a distinguishable feature representation. To evaluate the segmentation performance of the proposed method, we also collected a dataset of electron microscope images of membranous nephropathy. To the best of our knowledge, this is the largest image dataset for segmentation of glomerular basement membrane dense deposits. Experimental result shows that our model can accurately segment ordinary-sized dense deposits. Compared with state-of-the-art methods, our proposed method lower both false positive and false negative segmentation of small-sized protein sediments.