Dual discriminative local coding for tissue aging analysis
Graphical abstract
Introduction
With the longer life expectancy, aging-related health issues such as can- cer and neurodegenerative diseases are becoming more prevalent. Improving the quality of life at advanced age has become a critical problem in public health, and understanding the causal processes of aging is essential to fa- cilitate this development. However, currently there is no established theory of aging. While some researchers suggest aging is predetermined by genes, others believe it is the result of many lifelong influences (Sergiev et al., 2015). Aging studies have thus taken a multidisciplinary approach involving gene expression analysis, proteomics and metabolomics, pathway analysis, and image analysis (Wieser et al., 2011).
Image-based aging studies have focused on morphological age estimation from microscopy images (Wieser et al., 2011). Morphological age refers to the quantitative measure of age based on tissue morphology that can be visualized from microscopy images. It can represent the actual effect of aging on the particular individual, and individuals of the same chronological age can have different morphological ages. This information is important in aging research to discover the various factors causing different effects of aging on different people.
To quantify the morphological age of tissues, manual approaches have been conducted (Garigan et al., 2002, Herndon et al., 2002, Helfrich et al., 2007). These studies used manual scoring of morphological ages as the basis for further analysis and made interesting discoveries about the aging process. However, these approaches used different ways of scoring, without a standardization of the scoring process. Manual scoring can also be difficult to reproduce since it is hard to describe the visual criteria used for scoring. Especially if a more detailed quantification is expected rather than just differentiating between a few age categories, the manual approach would be subjective and unreliable. Techniques that can facilitate automatic quantification of tissue morphological age are highly desirable. In this paper, we present our approach to this problem.
Computer-aided approaches have been proposed for automated classifi- cation of tissue ages (Shamir et al., 2008b, Johnston et al., 2008, Shamir et al., 2009, Meng and Shyu, 2013, Coelho et al., 2013, Zhou et al., 2013, Pham, 2014). These approaches typically work by first extracting image features from various aspects then performing classification to separate the images into different age groups. For image feature description, texture fea- tures have mainly been used, including the traditional Haralick, Tamura and wavelet features (Shamir et al., 2008b, Johnston et al., 2008, Shamir et al., 2009, Zhou et al., 2013), more recent speeded-up robust features (SURF) (Coelho et al., 2013), and customized entropy features (Pham, 2014). Clas- sification is then performed using various algorithms, such as the nearest neighbor classifier (Shamir et al., 2008b, Pham, 2014), Fisher discriminant analysis (Johnston et al., 2008), and support vector machine (SVM) (Coelho et al., 2013, Zhou et al., 2013). Evaluation is then performed by comparing the derived age groups with the annotated chronological ages.
With the large variety of techniques, the age classification performance is however quite low with most of the studies reporting around 50% accuracy. Fig. 1 shows the example images from the terminal bulb aging database (Shamir et al., 2008a), which is commonly used in the aforementioned studies. It can be seen that there is a general trend of structural deterioration of muscle tissues with aging. However, it is hard to distinguish the degree of deterioration at different ages, and images of the same chronological age show large variations of visual characteristics. In addition, these approaches provide classification of chronological ages only, and do not support quantification of morphological ages. This could be due to the unavailability of ground truth for morphological ages, which is still a subjective measure in current aging research.
In a broader scope, tissue aging analysis is closely related to microscopy image classification, since they have the same overall framework that comprises feature extraction and classification. Depending on the focus of method design, existing studies in microscopy image classification can be categorized into two groups. The first group focuses on feature extraction, in which cus- tomized features are designed (Su et al., 2012, Sparks and Madabhushi, 2013, Peter et al., 2015, Xu et al., 2015, Jiang et al., 2015, Barker et al., 2016) or automated feature learning is conducted (Zhou et al., 2014, Otalora et al., 2015, BenTaieb et al., 2015, Wang et al., 2015). The second group focuses on the classifier design while standard and simple feature descriptors are used. In particular, multiple instance learning (Xu et al., 2014, Kandemir et al., 2014, Li et al., 2015) and sparse representation (Srinivas et al., 2014, Vu et al., 2015, Taalimi et al., 2015, Su et al., 2015) have been the major works in classifier design for microscopy image classification.
In this work we focus on classifier design and propose a sparse representation based method. Here we present a detailed review on sparse representation and its application in biomedical image analysis. Sparse representation performs classification based on the quality of reconstruction for the classification tasks (Wright et al., 2010). Typically the test data is sparsely reconstructed using reference dictionaries of different classes, and then classified to the class corresponding to the best reconstruction. Unlike SVM and other discriminative or generative classifiers, sparse representation is a non-parametric model and does not require training to model the separation between classes. Its data-driven approach is especially suitable when there is limited amount of training data to effectively exploit the discriminative structure in the feature space. Sparse representation has thus been a popular classifier for biomedical imaging applications (Xu et al., 2013, Weiss et al., 2013, Song et al., 2013, Srinivas et al., 2014, Wang et al., 2014).
On the other hand, the classification performance of sparse representation is entirely dependent on the quality of sparse reconstruction, and the classification objective is not associated with the reconstruction process. Since sparse reconstruction is optimized towards achieving good reconstruction of the test data given any reference dictionary, a better reconstruction could be obtained for the wrong class, leading to misclassification. Several ways have been proposed to improve the basic sparse representation model. For example, dictionary learning is used to generate a more discriminative reference dictionary from the image patches or feature vectors (Liu et al., 2011, Tong et al., 2013, Vu et al., 2015, Taalimi et al., 2015). Such methods are suitable for problems with large number of training data and small feature dimension. Another way is to restrict the selection of reference items that are used to reconstruct the test data. In locality-constrained linear coding (LLC) (Wang et al., 2010), the nearest neighbors of the test data are first identified and then linearly combined to obtain the sparse representation. LLC is highly efficient with its analytical solution and has been applied in various biomedical imaging studies (Zhang et al., 2013, Xing and Yang, 2013, Wu et al., 2014, Song et al., 2015b, Su et al., 2015). In addition, ensemble learning has been integrated with sparse representation, in which the sparse reconstruction is performed using subsets of reference dictionaries and the classification decisions are fused with boosting (Huang et al., 2014, Song et al., 2014), large margin optimization (Song et al., 2015a), or customized fusion weights (Song et al., 2015b). These approaches impose an additional layer of discrimination based on the reconstruction outputs and can effectively enhance the classification performance.
In this work, we present an automated method for quantifying the morphological age of tissues from microscopy images. Since currently there is no established rules to compute the morphological ages and automated approaches only support classification of chronological ages, our design is exploratory and we hypothesize that the tissue morphological age should be closely correlated with its chronological age. We thus propose to derive the morphological age by first classifying the chronological age of the tissue. The classification outputs give the degrees of correlation between the test image and various chronological ages. Then based on these correlation measures, the morphological age is computed. In summary, our method comprises three components: image feature extraction, chronological age classification, and morphological age quantification. Fig. 2 illustrates our method design.
Our main methodological contribution is the design of a new sparse representation model, namely the dual discriminative local coding (DDLC) method, which is used for the chronological age classification component. Briefly, we choose the LLC model as our base classifier, considering its computational efficiency and improved classification performance with the locality constraints. However, the discriminative capability of LLC can still be limited, because LLC does not involve learning-based optimization for classification and it derives the locality constraints simply based on Euclidean distances. Our DDLC method is designed to address these issues with dis- criminative distance computation and dual-level local coding. Our second contribution is that we have designed a new way of texture representation for the terminal bulb images. By combining multiple types of texture descriptors extracted at two levels, we obtain a highly descriptive image descriptor. Finally, we propose a method that quantifies the morphological ages based on the classification outputs of chronological ages. Evaluation is performed on the publicly available terminal bulb aging database (Shamir et al., 2008a), which is commonly used in studies for automated tissue aging analysis.
A preliminary version of this work has been published in the conference paper (Song et al., 2016). Different from the two-level LLC model proposed in the preliminary study, in this work, we have further incorporated discriminative learning into the sparse representation model and developed the new DDLC method. In addition, more thorough evaluation of the quantification results has been conducted. This paper provides substantially more detailed description of the problem domain and algorithm design.
Section snippets
Dual discriminative local coding
We first describe the preliminaries of the LLC algorithm and an LLC- based classification method. We then elaborate on the design motivation of our DDLC method, followed by presenting our design of dual-level local coding and discriminative distance computation.
Tissue aging analysis
In this section, we describe the dataset used for tissue aging studies, our design of image feature descriptors, and the way of using our dual discriminative local coding for morphological age quantification. The experimental setup is also described.
Chronological age classification
We first evaluate our method performance for classifying chronological ages, since existing approaches on automated tissue analysis mainly focus on such classification problems. Table 2 shows the classification results in comparison with the published results for the same terminal bulb database (Shamir et al., 2008b, Zhou et al., 2013, Coelho et al., 2013). We computed the mean and standard deviation of classification accuracies from the ten splits of five-fold cross validation, while the
Conclusions and future work
We present an automated method for quantifying the morphological ages of tissues. Our method consists of three steps. First, image features are extracted at two scales using multiple types of descriptors. Second, the chronological age of tissue is classified using our dual discriminative local coding (DDLC) model, which is designed based on the locality-constrained local coding method with additional discriminative distance learning and dual- level local coding to enhance the discriminative
Acknowledgment
This work was supported in part by Australian Research Council (ARC) grants.
References (60)
- et al.
Automated clas- sification of brain tumor type in whole-slide digital pathology images using local representative tiles
Med. Image Anal.
(2016) - et al.
Contour tracking in echocardiographic sequences via sparse representation and dictionary learning
Med. Image Anal.
(2014) - et al.
Locality- constrained subcluster representation ensemble for lung image classifica- tion
Med. Image Anal.
(2015) - et al.
Explicit shape descriptors: novel mor- phologic features for histopathology classification
Med. Image Anal.
(2013) - et al.
Seg- mentation of MR images via discriminative dictionary learning and sparse coding: application to hippocampus labeling
Neuroimage
(2013) - et al.
Integration of sparse multi-modality representation and anatomical con- straint for isointense infant brain MR image segmentation
Neuroimage
(2014) - et al.
Weakly supervised histopathology cancer image segmentation and classification
Med. Image Anal.
(2014) - et al.
Automatic Diagnosis of Ovarian Carcinomas Via Sparse Multiresolution Tissue Representation
(2015) - et al.
Deep Filter Banks For Texture Recognition and Segmentation
(2015) - et al.
Determining the subcellular location of new proteins from microscope images using local features
Bioinformatics
(2013)