Inflammatory Bowel Disease (ibd)

Inflammatory bowel disease (IBD) is a chronic inflammation of the gastrointestinal tract. IBD is classified into two major types, Crohn’s disease (CD) and Ulcerative colitis (UC). The prevalence of CD and UC is the highest in Europe with 322 and 505 per 100,000 persons respectively (Molodecky et al., 2012). Conventionally, the severity of IBD is diagnosed using histopathological examination performed by a trained pathologist. Morphological changes like crypt distortion, the presence of infiltrates in the lamina propria and erosion of the epithelial layer are used as inflammatory markers to predict the disease stage and plan a clinical therapy.

In the past decade, Label-free Multiphoton microscopy (MPM) has been recognized as a real-time invivo imaging technique for IBD. Its increased penetration depth, high spatial resolution and molecular specificity have accelerated the IBD diagnosis. MPM techniques like two-photon excited fluorescence (TPEF) and second harmonic generation (SHG) along with the coherent anti-stokes Raman scattering (CARS) can be used to visualize molecular changes associated with IBD (Schürmann et al., 2013).

Chernavskaia et al. used intensity related properties of CARS/TPEF/SHG and the crypt morphology to assign the histological index to a tissue section of an IBD patient. In their study, the mucosal and the crypt regions were annotated by a trained pathologist which is a labour- intensive and time-consuming task (Chernavskaia et al., 2016). Therefore, an automatic segmentation of the crypt and mucosa region using a multimodal image is a pre-requisite for estimating the histological index associated with the different IBD stages.

Nevertheless, automatic segmentation of the crypt and mucosa region is a very challenging task due to several reasons. First, the crypt morphology changes between patients of different disease activity. The crypt structure is distorted for patients with higher IBD stage. Second, the crypts are located within the mucosa region and therefore the two regions overlap which makes the classification even more challenging. Third, identifying clear boundaries of the crypt structure is difficult as the crypts are very closely located to each other. Lastly, there is a limited availability of annotated medical data which captures various tissue structures of an IBD patient. Therefore, segmentation of these regions by image processing and classical machine learning techniques is inefficient.

Semantic segmentation using Deep Convolutional Neural Network (DCNN) has achieved successful results in the past. Deep neural networks like the U-Net, SegNet have been used for biomedical image segmentation and are the benchmark for pixel-wise segmentation. In this paper, we propose an automatic segmentation of multimodal images into four regions using a DCNN. Further, we compare the segmentation results obtained by DCNN with classical machine learning approach.

The paper is organized as follows, in section (2) we introduce the previous work related to gland segmentation using histology images, in section (3) we introduce our multimodal image dataset and our segmentation workflow. This is followed by evaluation metrics and results in section (4). We discuss and conclude our work in section (5).