Home Engineering Cellular Image Classification
Directions and Future Work
In recent years, although considerable progress has been made, research in HEp-2 cell image analysis is still in its early stage and there is great potential for improvement. For future work, we suggest some long-term research goals for staining pattern classification.
Firstly, designing corresponding features is a prerequisite for image classification. It is not easy to find effective features for staining patterns of HEp-2 cells as the distinction between different patterns is subtle. It is even difficult for human beings to separate them. In this book, we have proposed three kinds of features to characterizing the staining patterns. However, there are still lots of developed features existed such as BRISK , ORB , etc. Moreover, the mid-level features which are transformed from low-level features should be researched more deeply. On the other hand, multiple features can be heuristically combined into one high dimensional feature to complement each other. In addition, feature selection is a procedure to select a subset of the most relevant features from the input data which can describe the input data efficiently and provide good prediction results. It can choose the most discriminative features, reduce the measurement and storage requirements, reduce computational complexity and training time, defy the curse of dimensionality to improve classification performance . There are plenty of feature selection methods available in literatures, e.g., filter methods, wrapper methods, embedded methods . We will further study on extracting and selecting effective and reasonable features to satisfy the need of staining pattern classification. Furthermore, we plan to explore the feature extraction and classification for general images.
Secondly, deep learning, also called hierarchical learning, which is a branch of machine learning, has become a hot topic in recent years. Until now, most machine learning methods had exploited shallow structured architectures such as SVM, logistic regression, multilayer perceptrons (MLPs), etc. However, human information processing mechanisms suggest to construct deep architectures to extract complex structure and build internal representation from rich sensory inputs . It is reasonable that the state-of-the-art can be improved if appropriate deep learning algorithms can be developed. There are plenty of deep architectures, such as Deep Neural Networks (DNNs), Deep Belief Network (DBN), etc. They have been successfully applied to speech recognition [6, 7] and computer vision [8, 9]. However, they have not been developed for staining pattern classification. In the future, we will study deeply on this strong learning method and improve it for staining pattern classification.
Last but not least, we will research on the features and classifiers for fluorescent intensity classification. Some researchers have investigated into the fluorescent intensity classification. For example, Soda and Iannello  extract a set of statistical features, which is based on the first-order and the second-order gray-level histograms, from the whole image. For intensity classification, a multiple expert approach based on three classifiers is adopted. Each classifier specialized in identifying one of three input classes (i.e., positive, intermediate and negative). Rigon et al.  present a comprehensive system supporting the ANA test. It includes two systems respectively classifying the fluorescence intensity and staining pattern. For fluorescence intensity, they uses the multi expert/module system and propose two rules to provide the final classification. However, their researches are evaluated by using private datasets and experimental protocols. Next, we are exploring features and classifiers to improve the performance of fluorescence intensity classification. Our final goal is to build an automatic system to support the doctors’ diagnosis which should contain two major tasks, i.e., fluorescent intensity classification and positive staining pattern classification.
|< Prev||CONTENTS||Next >|