False positive reduction in urinary particle recognition

Abstract

The identification of non-cell objects in biological images is not a trivial task largely due to the difficulty in describing their characteristics in recognition systems. In order to better reduce the false positive rate caused by the presence of non-cell particles, we propose a novel approach using a local jet context features scheme combined with a two-tier object classification system. The newly proposed feature scheme, namely local jet context feature, integrates part of global features with the “local jet” features. The scheme aims to effectively describe the particle characteristics that are invariant to shift and rotation, and hence help to retain the critical shape information. The proposed two-tier particle classification strategy consists of a pre-recognition stage first and later a further filtering phase. Using the local jet context features coupled with a multi-class SVM classifier, the pre-recognition stage intends to assign the particles to their corresponding classes as many as possible. To further reduce the false positive particles, next a decision tree classifier based on shape-centered features is applied. Our experimental study shows that through the proposed two-tier classification strategy, we are able to achieve 85% of identification accuracy and 80% of F₁ value in urinary particle recognition. The experiment results demonstrate that the proposed local jet context features are capable to discriminate particles in terms of shape and texture characteristics. Overall, the two-tier classification stage is found to be effective in reducing the false positive rate caused by non-cell particles.

Introduction

The urinary sediment analysis of biological particles in microscopic images has been a subject of research for many years in pathology (Lamchiagdhase et al., 2005, Linko et al., 2006). A correct urinalysis result offers a direct indication of the patient’s renal and genitourinary system and helps to monitor other body systems (Chien et al., 2007). Urinary sediment analysis is usually performed by trained technicians who count the number of each kind of particles using a microscope focusing on the ground slide with urinary sample or through a monitor screen displaying the biological particles that are digitally captured from the urinary sample. Although the procedures for manual analysis have already been standardized, the traditional microscopy approach in urinary sediment analysis is laborious and tedious which often produces subjective and imprecise results (Lamchiagdhase et al., 2005). Hence, there is an increasing demand for an automated system that can detect and count the suspicious objects in the urinary images and further classify them into the correct classes of interests.

In the typical procedures of automated urinalysis, detection and classification are both challenging tasks. Basically, this is due to the complicated characteristics of urinary images:

• The image backgrounds can be noisy and non-uniform in illumination.

• The objects of interest exist in diversified forms. Even for the items belong to the same class, they can vary greatly in shape and texture.

• The presence of spots caused by multilayer urinal sample slides makes the problem more complex. Some particles can be presented in poor resolution and their boundaries are not sharp enough to be readily extracted.

• The existence of non-cell particles in different texture and shape will surely jeopardize the recognition results.

As a result, some non-cell particles will be inevitably labeled as cells, e.g. red blood cell (RBC) or white blood cell (WBC), which directly leads to in a high false positive rate.

Recently, it has witnessed an increasingly attention amongst researchers and practitioners in handling image classification in cell biology (Auer, Peng, & Singh, 2007). Although existing algorithms are highly sensitive in cell classification, they mainly focus on the identification of biological cells only, rather than the process of identifying non-cell particles (such as impurity and poor focused regions). There are two main reasons why non-cell objects exist in the images. One is due to the image quality problem from poor focusing regions which are located in various places of urinary images. The other reason is because of the random occurrence of impurities. In some urinary images, the impurities are rare; while in some other cases, the impurities appear continuously along a series of images. Inevitably, some non-cell particles are likely to be assigned to one or the other cell group. This will increase the number count in different cell groups of interest. As a result, technicians have to manually examine the specific cases with relatively a large number of cells from certain classes. Therefore, it is highly desirable to lower the false positive rate caused by non-cell particles as much as possible, while retaining the true positives, i.e. RBCs and WBCs.

In literature, several strategies tackling cell image recognition problem have been reported. Here, we only focus on the approaches adopted in feature extraction scheme, which is well recognized as one of the most important stages in pattern recognition, and summarize them based on their original intents of how to describe the object characteristics.

The first prevailing approach is established based on global features. Global features are intended to describe the regions of interest in a macroscopical way. Such features can refer to the statistics of pixel intensities, e.g. mean, standard deviation and correlation between adjacent pixels, or various shape metrics, e.g. area, perimeter and elongation, or edge information, e.g. skewness and curvature, or colorimetric features, etc. Various feature selection techniques will be applied in order to generate an effective feature set usually in a lower dimension (Liyan, Senmiao, Guangyuan, Lingyan, & Yongli, 2007). Using the reduced feature set as variable inputs, classifiers are induced based on different machine learning algorithms, e.g. k-nearest neighbor, artificial neural network, Bayesian classifier and support vector machines, and labeled learning examples to predict any unseen objects. Often, the identification and screening of these features are domain based and highly rely on the knowledge of a specific application, therefore, rule based methods are also proposed (Chen et al., 2006, Lin et al., 2005, Zhen et al., 2000). However, the limitation of this knowledge based rule approach is well understood since it is often intended for certain types of cells. As for other types of particles, the feature discrimination procedure would not be valid (Dahmen et al., 2000, Liyan et al., 2007). This also jeopardizes its robustness in different image transformations (e.g. orientation, scale, translation and symmetrical illumination) that appear in many applications.

The second major effort emphasizes on local appearance models. It has provided another class of methods in terms of the source of features. From the microcosmic point of view, local features are mainly extracted based on the appearance of particularly interesting points or regions. The local descriptors propose extensive local features which are invariant to various transformations by encoding the numeric information in their neighborhood (Krueger and Phillips, 1989, Mikolajczyk and Schmid, 2005, Schmid and Mohr, 1997). Due to the well-known challenge of high dimensionality in feature space with respect to various machine learning algorithms, feature reduction is conducted, e.g. principal component analysis and linear discriminant analysis. Kolsch et al. introduced squared sub-images with a high local variance of gray values as local features (Kolsch, Keysers, Ney, & Paredes, 2004). Distance metrics and probability models were also proposed in image classification. The experimental study stresses that the decomposition of the training images into sets of local features with respect to each class of interest is the most important aspect in the framework of image classification. Maree & Wehenkel proposed a generic approach for image classification without segmentation (Maree, Geurts, & Wehenkel, 2007). A large set of patches were randomly extracted and their raw pixel values were converted to high-dimensional feature vectors. Next, these features were adopted in the extremely randomized trees to build a subwindow classification model. Although the results were however not as good as the best results for one specific dataset, they illustrated the potential of their generic image classification method. Ranzato et al. presented an automatic system to classify 12 object categories in urinary images (Ranzato et al., 2007). Without any segmentation, the differential invariants of brightness at multiple scales were described by an average operator. With these invariant features, a classifier obtained from a mixture-of-Gaussian generative model achieved a 93.2% accuracy. It has come to our attention if the focusing objects from different classes in various shapes bear similar textures, these methods may not be able to perform well. This is highly because the average operator and the subwindows features do not provide sufficient information about size, shape and density distribution of the particles.

The combination of global and local features has emerged as another stream in feature description. Both global and local features have their own merits in representing different kinds of cells. The utilization of a joint feature set may be more effective in classifying certain types of cells than others. How to integrate these two classes of features is indeed a research problem. In literature, studies have been reported to generate feature vectors by joining global and local features together before supplying to learning algorithms. Wang et al. built feature vectors with 145 features including shape, size and Gabor transformed features (Wang, Zhou, King, & Wong, 2007). Lucas et al. derived features from the sub-images and texture-related features. Feature selection techniques are applied to filter out the less significant features (Lucas, Raul, Bruno, Anne, & Marcilio, 2007). A notable problem in their approach is that the normalization operator may severely jeopardize the contribution of certain features in classification, largely due to their different orders of magnitude when compared to each other. One possible solution is to arrange these features in a hierarchical manner using different learners. Zhen et al. built artificial neural networks to classify normal substance, and then fuzzy rules were generated according to the global features chosen (Zhen et al., 2000). However, their experiments have been limited to two substances only, the white blood cells and the red blood cells. Further exploration and validation are needed.

In our study, the non-cell object class possesses many more samples than the classes of biological cells. Our purpose is to reduce the false positives, e.g. the impurity being regarded as a red blood cell or other type of cells, caused by non-cell objects as many as possible, meanwhile not jeopardizing, if not improving, the true positive rate. In this paper, we mainly focus on feature extraction scheme which is often regarded as one of the most important stages in particle classification problem as well as the strategy concerned about how to combine the global and local features.

Fig. 1 shows an overview of the proposed approach. First, we employ a global thresholding approach of local difference to detect the particles. Next, a two-step procedure for particle segmentation has been adopted. Once the individual particle is segmented, feature extraction is carried out. Local descriptors are a powerful technique for small object representation, such as SIFT, steerable filters and local jet. Based on the local descriptor “local jet”, we present a new feature descriptor namely the local jet context (LJC) features to describe the characteristics of the detected regions. Some useful shape-based features are also introduced according to the biological knowledge. After this, a two-tier object classification system is introduced which consists of a pre-recognition stage first and later a further filtering stage. Using the local jet context features coupled with multi-class SVM classifiers, the pre-recognition stage aims to assign the particles to their corresponding classes as many as possible. To further remove the false positive particles, next a decision tree based on shape-centered feature is applied. Finally, the system outputs the identification results.

The rest of this paper is organized as follows. We introduce a two-step procedure for particle segmentation in Section 2. Section 3 mainly focuses on feature extraction especially our proposed feature scheme. The experiment setup is explained in Section 4. Section 5 reports our experimental results and analysis. Section 6 concludes with some future work.