A computer aided diagnosis system for lung nodules detection in postero anterior chest radiographs

This thesis describes a Computer Aided System aimed at lung nodules detection. The fully automatized method developed to search for nodules is composed by four steps. They are the segmentation of the lung field, the enhancement of the image, the extraction of the candidate regions, and the selection between them of the regions with the highest chance to be True Positives. The steps of segmentation, enhancement and candidates extraction are based on multi-scale analysis. The common assumption underlying their development is that the signal representing the details to be detected by each of them (lung borders or nodule regions) is composed by a mixture of more simple signals belonging to different scales and level of details. The last step of candidate region classification is the most complicate; its 8 task is to discern among a high number of candidate regions, the few True Positives. To this aim several features and different classifiers have been investigated. In Chapter 1 the segmentation algorithm is described; the algorithm has been tested on the images of two different databases, the JSRT and the Niguarda database, both described in the next section, for a total of 409 images. We compared the results obtained with another method presented in the literature and described by Ginneken, in [85], as the one obtaining the best performance at the state of the art; it has been tested on the same images of the JSRT database. No errors have been detected in the results obtained by our method, meanwhile the one previously mentioned produced an overall number of error equal to 50. Also the results obtained on the images of the Niguarda database confirmed the efficacy of the system realized, allowing us to say that this is the best method presented so far in the literature. This sentence is based also on the fact that this is the only system tested on such an amount of images, and they are belonging to two different databases. Chapter 2 is aimed at the description of the multi-scale enhancement and the extraction methods. The enhancement allows to produce an image where the “conspicuity” of nodules is increased, so that nodules of different sizes and located in parts of the lungs characterized by completely different anatomic noise are more visible. Based on the same assumption the candidates extraction procedure, described in the same chapter, employs a multi-scale method to detect all the nodules of different sizes. Also this step has been compared with two methods ([8] and [1]) described in the literature and tested on the same images. Our implementation of the first one of them ([8]) produced really poor results; the second one obtained a sensitivity ratio (See Appendix C for its definition) equal to 86%. The considerably better performance of our method is proved by the fact that the sensitivity ratio we obtained is much higher (it is equal to 97%) and also the number of False positives detected is much less. The experiments aimed at the classification of the candidates are described in chapter 3; both a rule based technique and 2 learning systems, the Multi Layer Perceptron (MLP) and the Support Vector Machine (SVM), have been investigated. Their input is a set of 16 features. The rule based system obtained the best performance: the cardinality of the set of candidates left is highly reduced without lowering the sensitivity of the system, since no True Positive region is lost. It can be added that this performance is much better than the one of the system used by Ginneken and Schilam in [1], since its sensitivity is lower (equal to 77%) and the number of False Positive left is comparable. The drawback of a rule based system is the need of setting the 9 thresholds used by the rules; since they are experimentally set the system is dependent on the images used to develop it. Therefore it may happen that, on different databases, the performance could not be so good. The result of the MLPs and of the SVMs are described in detail and the ROC analysis is also reported, regarding the experiments performed with the SVMs. Furthermore, the attempt to improve the performance of the classification leaded to other experiments employing SVMs trained with more complicate feature sets. The results obtained, since not better than the previous, showed the need of a proper selection of the features. Future works will then be focused at testing other sets of features, and their combination obtained by means of proper feature selection techniques.