Tubule segmentation of fluorescence microscopy images based on convolutional neural networks with inhomogeneity correction

Fluorescence microscopy has become a widely used tool for studying various biological structures of in vivo tissue or cells. However, quantitative analysis of these biological structures remains a challenge due to their complexity which is exacerbated by distortions caused by lens aberrations and light scattering. Moreover, manual quantification of such image volumes is an intractable and error-prone process, making the need for automated image analysis methods crucial. This paper describes a segmentation method for tubular structures in fluorescence microscopy images using convolutional neural networks with data augmentation and inhomogeneity correction. The segmentation results of the proposed method are visually and numerically compared with other microscopy segmentation methods. Experimental results indicate that the proposed method has better performance with correctly segmenting and identifying multiple tubular structures compared to other methods. Introduction Recent advances in fluorescence microscopy imaging, especially two-photon microscopy, have enabled the imaging of cellular and subcellular structures of living tissue [1, 2, 3]. This has resulted in the generation of large datasets of 3D microscopy image volumes, which in turn need automatic image segmentation techniques for quantification. However, the quantitative analyses of these datasets still pose a challenge due to light scattering, distortion created by lens aberrations in different directions, and the complexity of biological structures [4]. The end result is blurry image volumes with poor edge details that become worse in deeper tissue depths. There have been various techniques developed for segmentation of microscopy images. One widely used class of methods is based on the active contours technique which minimizes an energy functional to fit contours to objects of interest [5, 6]. Early version of active contours [5] generally produced poor segmentation results since the segmentation results are noise sensitive and initial contour dependent. An external energy term which convolves a controllable vector field kernel with an image edge map was presented in [7] to address the noise sensitive problem. Similarly, the Poisson inverse gradient was introduced to determine initial contours locations to segment microscopy images in [8]. Moreover, active contours methodology has been integrated with a region-based segmentation method that poses the segmentation problem as an energy equilibrium problem between foreground and background regions [9]. In addition, this region-based active contours technique was extended to fully utilize 3D information to identify foreground and background voxels [10]. More recently, this 3D region-based active contours was combined with 3D inhomogeneity correction to provide better segmentation since this technique takes into consideration inhomogeneities in volume intensity [11]. Additionally, a new segmentation method known as Squassh that couples image restoration and segmentation using a generalized linear model and Bergman divergence was introduced [12], whereas a method that combined with detecting primitives based on nuclei boundaries and identifying nuclei region using region growing was demonstrated in [13]. Alternatively, combination of multiresolution, multiscale, and region growing methods using random seeds to perform multidimensional segmentation was described in [14]. As indicated above, florescence image segmentation still remains a challenge problem. Tubule, a biological structure with a tubular shape, segmentation is even more challenging since tubular shape and orientation is varied without known patterns. Also, since typical tubular structures have hollow shapes with unclear boundaries, traditional energy minimization based methods such as active contours have failed segmenting tubular structures [15]. There has been some work particularly focusing on tubular structure segmentation. A minimal path based approach was described in [16, 17] where tubule shape is modeled as the envelope of a family of spheres (3D) or disk (2D). Similarly, a new approach for 3D human vessels segmentation and quantification using 3D cylindrical parametric intensity model was demonstrated in [18]. Also, multiple tubule segmentation technique that combined with level set methods and the geodesic distance transform was introduced in [19]. More recently, one method used to segment tubular structures was delineating tubule boundaries followed by ellipse fitting to close the boundaries while considering intensity inhomogeneity [15]. Another method known as Jelly filling [20] utilized adaptive thresholding, component analysis, and 3D consistency to achieve segmentation, whereas a method for tubule boundary segmentation used steerable filters to generate potential seeds from which to grow tubule boundaries followed by tubule/lumen separation and 3D propagation to generate segmented tubules in 3D [21]. Previous methods, however, focused on segmenting boundar X iv :1 80 2. 03 54 2v 1 [ cs .C V ] 1 0 Fe b 20 18 aries of tubule membrane. Since some tubule membranes are not clearly delineated in fluorescence microscopy image volume, finding tubule boundaries may not always result in identifying individual tubule regions. Convolutional neural network (CNN) has been used to address segmentation problems in biomedical imaging [22]. The fully convolutional network [23] introduced an encoder-decoder architecture for semantic segmentation. U-Net [24] is a 2D CNN based method utilizing this encoder-decoder architecture with connecting intermediate stages of downsampling and upsampling to preserve information. U-Net can be used segment complex biological structure in microscopy images. Similarly, in [25] a U-Net trained on cell objects and contours was used to identify tubular structures. Additionally, a multiple input and multiple output structure based on a CNN for cell segmentation in fluorescence microscopy images was demonstrated in [26]. Also, a nuclei segmentation method that combined with a 2D CNN and a 3D refinement process was introduced in [27]. In this paper, we present a method for segmenting and identifying individual tubular structure based on a combination of intensity inhomogeneity correction, data augmentation, followed by a CNN architecture. Our proposed method is evaluated at objectlevel metrics as well as pixel-level metrics using manually annotated groundtruth images of real fluorescence microscopy data. Our datasets are comprised of images of a rat kidney labeled with a phalloidin which labels filamentous actin collected using twophoton microscopy. A typical dataset we use in our studies consists of two tissue structures, the base membrane of the tubular structures and the brush border which is generally located interior to proximal tubules. Our goal here is to segment individual tubules enclosed by their membranes.