Medical Image Segmentation (mis)

Medical Image Segmentation (MIS) has been applied to numerous applications like delineating tissue structures, cell counting, lesion and tumour tracking etc. Normally, the approach for MIS can be categorized into three types. First, segmentation using classical image processing techniques like thresholding, morphological operations, and watershed transform. Second, to train a classification model based on hand-crafted features like statistical features, grey level co-occurrence matrix, local binary pattern etc. The third approach is segmentation using high-level features obtained by a DCNN. Wu et al. used classical image processing algorithms including thresholding and seeded region growing for segmentation of the human intestinal glands. However, this method considered a prior knowledge of the morphological structures in the gland and was evaluated qualitatively (Wu et al., 2005).

In another approach by Peng et al., a k-means clustering and morphological operations were used to segment the prostate glandular structures. Based on these structures a linear classifier to distinguish normal and malignant glands was built (Peng et al., 2010). Feature extraction and selection has been widely used in application areas like biomedicine, image analysis, biometric authentication etc. In the contribution of Farzam et al. and Doyle et al., texture, shape and graph-based features were extracted and a linear classifier was built to distinguish different pathological tissue sections of the prostate cancer patients (Farzam et al., 2007) (Doyle et al., 2007). In the work presented by Naik et al., a Bayesian classifier was used to classify between lumen, stroma and nuclei.

The true lumen areas were identified by applying size and structure constraints. A level set curve was initialized using the true lumen area and was evolved until the interior boundary of the nuclei. Morphological features were calculated from the boundaries followed by a manifold learning scheme to classify cancer grades based on the reduced features (Naik et al., 2008). By the previous methods, regularly shaped gland structures were efficiently segmented. However, due to various sample preparation factors, the gland structures show variation and to segment irregularly shaped gland structures is a challenge. To alleviate this problem, Gunduz-Demir et al. proposed an object-graph based approach that relies on decomposing the image into objects. Their approach used a three-step region growing algorithm, followed by boundary detection and false region elimination (Gunduz-Demir et al., 2009). In another work by Sirinukunwattana et al., a Random Polygons Model to segment glandular structure in human colon tissue was formulated. The glandular structures were modelled as polygons whose vertices were located on the epithelial border nuclei. At first, the glandular probability map was generated using super-pixel texture features, this was followed by identifying nuclei vertices and constructing random polygons from seed areas. False positive polygons were removed by post-processing procedures (Sirinukunwattana et al., 2015).

Nowadays, deep learning techniques have achieved promising results in MIS. The most relevant DCNN like AlexNet (Krizhevsky et al., 2012), VGGNet (Simonyan et al., 2014), GoogLeNet (Szegedy et al., 2014), U-Net (Ronneberger et al., 2015) and SegNet (Badrinarayanan et al., 2015) have achieved promising results in the past. The recent MICCAI 2015 Gland Segmentation Challenge presented several innovative algorithms for segmentation of the colon gland in the histology images (Sirinukunwattana et al., 2016). Chen et al. achieved state-of-the-art performance on the Warwick-QU colon adenocarcinoma dataset by integrating multi-level feature representation with Fully Convolutional Network (FCN) (Chen et al., 2015). Whereas, Kainz et al. used two DCNN that were inspired by the LeNet-5 architecture (LeCun et al., 1998) (Kainz et al., 2015). The first DCNN was used to separate the closely situated gland structures and the second DCNN was used to distinguish gland and non-gland regions (Kainz et al., 2015). In Awan et al., DCNN was used to delineate the gland boundaries and based on the glandular shape, a two-class and three-class classification model for colorectal adenocarcinoma using histology image were designed (Awan et al., 2017). In this paper, we intend to use SegNet (Badrinarayanan et al., 2015) for the segmentation of multimodal images into four distinct regions. Our method is different in the following ways:

• The above methods have been tested on H&E (Hematoxylin and Eosin) stained image which is an invasive technique and needs long sample preparation time. However, non-linear multimodal imaging can be used as an in-vivo technique and its automatic tissue classification can provide a real-time histological index.
• A rotation invariant data augmentation of the multimodal image is used to train the SegNet model.