The second test using a scanning NMR phantom: the black-circled area is a blastoma 

Context in source publication

Context 1

... mapping function retrieves 1 for the element a i corresponding to i* (i.e. the winning neuron), while the other outputs are set to 0. A SOM network has two layers, X (input layer) and Y (output layer) representing the classification results. There are two type of links: from X layer to Y layer (defined as W) and among units of Y layer, known as lateral links; they are organized in a two-dimensional grid-like schema (each unit is linked only to its “neighbors”). Considering an input X, the algorithm calculates euclidean distance between X and each classification unit. Then, the algorithm chooses the unit with minimal distance as the best-representing class. During the training phase, the SOM repeatedly analyze the inputs and, for each new example, the winning unit is found; thus, only the weights which link the winning unit with its neighbors are updated according to a specific topology. A weight variation increases the membership of input X to the calculated winning class (Yv). At the end of training phase, SOM neurons are disposed according to a well-defined spatial order; it is based on determination of weights of connections between classification units (Fig. 4). Therefore, it is possible to denote a spatial disposition of SOM neurons as similar as possible with the input space, so that inputs with similar features are rightly related to the same classification unit. In the analyzed case of study, it has been necessary to pursue a trade off between performances of trained SOM and its complexity. In fact, classification performances of SOM increase if number of SOM neurons is comparable to the dimension of input space. On the other hand, a high number of neurons compromises the quick functioning of classifier in a real-time application. Therefore, a number of neurons equal to considered cerebral tissues has been used (i.e. 12 neurons). It is a restricted number if compared with the dimension of input space, but adequate to assure improved performances of a SOM-based segmentation (Kohonen, 1997). Image used in Fuzzy-SOM test (Fig. 5) has been downloaded from the same BrainWeb site (BrainWeb, 2006). In this image, it is possible to detect some damaged cerebral areas (i.e. areas interested by sclerosis), which are highlighted with white circles. The proposed Fuzzy-SOM hybrid approach has been applied to this test image, in order to verify if the algorithm is able to distinguish damaged areas from the other biological tissues. Obtained results are reproduced in Fig. 6. In order to improve the comparison between observed and simulated data, only damaged areas are depicted into observed test image. Moreover, for a first evaluation of proposed heuristic approach, Fuzzy-SOM structure classify the presence or absence of pathological damages. From Fig. 6, it is possible to verify a good classification performance of proposed Fuzzy-SOM approach. Areas interested by sclerosis are all recognized; some misclassification are present, but they are negligible to diagnostic aims. In order to give a greater practical utility to proposed approach, a set of test images has been collected by using real NMR phantom of patients having cerebral damages of various nature and various extension. The new test set has been furnished by the Neuro-radiology Division of “Bianchi-Melacrino-Morelli” Hospital of Reggio Calabria, Italy. Since the hospital has not a digital version of database, phantoms have been digitalized by means of a specific scanner, having a double-illumination and a vertical-scanning: it is in equipment to the “Bone Marrow Transplant Center” of Reggio Calabria, Italy. Naturally, the acquisition process and the degradation of some phantoms involves a remarkable informative loss in terms of resolution and details of considered test imageries. The collected test set has been submitted as inputs to the Fuzzy-SOM network, obtaining encouraging. In the following subsection, two examples are going to be described. The first examined case concerns a NMR image with the following medical report: in cortical area with a dural basis, a most probably a secondary damage . The correspondent NMR has been passed to the trained Fuzzy- SOM classifier in order to obtain a segmented image and verify the abilities of heuristic classification, i.e. abilities of clustering of different tissues. Fig. 7 shows the classification result. It is evident how segmentation process carried out by proposed Fuzzy-SOM method has been able to recognize the damage, by evidencing interested area with a dark-green hue and distinguishing it into the context of the whole NMR phantom (Fig. 7). A similar process has been carried out with a second NMR phantom. In this case, medical report is: presence of a blastoma with an oval extension in cortical area . Fig. 8 shows result of classification carried out by Fuzzy-SOM network. Black-circled area depicts the blastoma lesion; it has been coloured with a dark-blue hue like the skull tissue because of NMR phantom degradation. In spite of this, result is a valid medical support, because blastoma has been distinguished from the rest of inner-cerebral tissues, whereas a doctor can excluding the black-circled area is skull for its same location. Other real cases have been considered in order to validate the efficiency and performances of proposed approach. The new test set reveals that segmentation process is very able to detect different kinds of tissues in NMR images, with a particular interest for pathological events. Moreover, by means of NMR images furnished by “Bianchi-Melacrino-Morelli” hospital, it has been proved that Fuzzy-SOM network is able to detect cerebral areas interested by any pathologies even if the same pathologies are not represented in training pattern set. In this paper the problem of pattern recognition in remote sensing imagery with medical applications has been examined. In particular, the attention has been focused on cerebral pathologies detection from NMR phantoms. Since they are determined by about a ten of physical parameters changing with impulse sequence at used radiofrequency, imagery interpretation is more complex than other non-invasive medical diagnostic methods. In order to solve this inverse problem, an unsupervised heuristic approach has been proposed for pattern recognition; it is based on Fuzzy clustering and Self-Organizing Neural Networks. Fuzzy C-means has been used in order to retrieve cluster centers of various classes (i.e. kind of cerebral tissues showed in a MRI). In order to improve classification performances, a SOM network has been trained by a suitable data set, using the cluster centers obtained by Fuzzy C-Means as initial values of SOM’s cluster centers. Fuzzy-SOM approach has been subsequently tested on a MRI available at BrainWeb site (BrainWeb, 2006) and on a set of NMR phantom images kindly provided by “Bianchi-Melacrino-Morelli” Hospital of Reggio Calabria, Italy. By analyzing retrieved results, it is possible to affirm that proposed network can highlight different type of pathologies in an appreciable way, even if the actual pathology were not represented into the training set. In conclusion, let us remark the efficiency and reliability of proposed approach in imagery segmentation and classification; thanks to its flexibility and adaptation abilities, the Fuzzy-SOM network has been able to establish the complex linkages which exist into the input space, generally joining similar inputs to a same output class. Authors are very grateful to staff of Neuro-radiology Division of “Bianchi Melacrino Morelli” Hospital (Reggio Calabria, Italy), and to staff of “Bone Marrow Transplant Center” (Reggio Calabria, Italy) for the useful and very experienced ...