Computer-Aided Diagnosis

Recently, Computer Aided Diagnosis (CAD) has become a part of the routine clinical works. With CAD, radiologists use the computer output as a second opinion to make the final decisions. The performance by CAD does not have to be comparable to or better than that by physicians, but should be complementary to that by physicians. In practice, a large number of CAD systems have been employed for assisting physicians in the early detection of breast cancers on mammograms.

Worldwide, there is a strong demand for IIF image analysis due to its effectiveness and high-quality. However, the main drawback of the strategy is the human subjective evaluation since the classification always depends on highly qualified experts’ skill. Results are significantly influenced by one’s qualification and reading systems, causing high-rate intra- and inter-laboratory variance. The low level of standardization limits the communications between the clinic units and reproducibility of IIF readings. As the demand for applying IIF method in diagnosis of autoimmune diseases is increasing, lacking of resources and adequately trained personnel is becoming another bottleneck [6, 17].

To address aforementioned issues, CAD systems are desired for automatically supporting the IIF diagnosis in many ways. They can augment the physicians’ capabilities, reduce the time and improve the reliability. They free the physicians from mass screening tasks and enable them to concern only about the most involved cases. Moreover, the CAD systems can be used as an education tool to train specialized medical personnel.

Several approaches have been proposed in the recent research for all the major stages of the IIF diagnostic procedure. The main technologies have been investigated in the CAD systems are automated preparation of slides with robotic devices [18], image acquisition [8, 21], image segmentation [15, 16], mitotic cell recognition [4], fluorescence intensity classification [18, 19] and staining pattern classification [6, 7, 20]. Till now, the highest level of automation in the CAD systems for ANA test is the preparation of slides by robotic devices conducting dilution, dispensation and washing operations [4]. In image acquisition stage, images are taken using a fluorescence microscope, equipped with both a mercury vapour lamp and a digital camera to auto-focus; then the digital images are displayed on a computer screen. Images are analyzed and processed to be more suitable to following classification. Fluorescence intensity is classified into negative group and positive group with intensity level based on the intensity-related features. The positive group is further classified into several main staining pattern groups based on the pattern-related features.

Sample images from ICPR 2012 contest dataset and ICIP 2013 contest training dataset respectively

Fig. 1.5 Sample images from ICPR 2012 contest dataset and ICIP 2013 contest training dataset respectively. The rows named “positive” are the patterns with positive intensity, while the “intermediate” rows are the patterns with intermediate intensity

While all aspects of the CAD systems contribute to the automation of IIF procedure in one way or another, staining pattern classification is proven to be the most challenging task in the research community due to large intra-class variation and small between-class variation regardless of its importance. To reduce the variability of multiple readings, the levels of fluorescent intensity are always generally classified into three levels named negative, intermediate (with intensity of 1+) and positive (with intensity of 2+ or more), where intermediate and positive is belong to the positive group whose pattern need to be further identified. In the following of this thesis, when we say “positive staining patterns”, we refer to staining patterns with non-negative fluorescence intensity, which includes positive and intermediate level. As shown in Fig. 1.5, the cells with intermediate and positive intensity which are from the same pattern categories have large variations. Particularly, the cells with intermediate intensity can not be seen clearly. Meanwhile, some categories share similar shape and texture. Image representation is crucial for HEp-2 cells analysis. Compared with the signal in general image classification, HEp-2 cells do not contain abundant structural information. In addition, the features between various HEp-2 cells are much more similar than those between different objects or natural scene images. In this thesis, we investigate into the feature extraction and machine learning methods for automatic staining pattern classification of HEp-2 cells.

< Prev   CONTENTS   Source   Next >