G. Coppini

Neural networks for computer-aided diagnosis: detection of lung nodules in chest radiograms

The paper describes a neural-network-based system for the computer aided detection of lung nodules in chest radiograms. Our approach is based on multiscale processing and artificial neural networks (ANNs). The problem of nodule detection is faced by using a two-stage architecture including: 1) an attention focusing subsystem that processes whole radiographs to locate possible nodular regions ensuring high sensitivity; 2) a validation subsystem that processes regions of interest to evaluate the likelihood of the presence of a nodule, so as to reduce false alarms and increase detection specificity. Biologically inspired filters (both LoG and Gabor kernels) are used to enhance salient image features. ANNs of the feedforward type are employed, which allow an efficient use of a priori knowledge about the shape of nodules, and the background structure. The images from the public JSRT database, including 247 radiograms, were used to build and test the system. We performed a further test by using a second private database with 65 radiograms collected and annotated at the Radiology Department of the University of Florence. Both data sets include nodule and nonnodule radiographs. The use of a public data set along with independent testing with a different image set makes the comparison with other systems easier and allows a deeper understanding of system behavior. Experimental results are described by ROC/FROC analysis. For the JSRT database, we observed that by varying sensitivity from 60 to 75% the number of false alarms per image lies in the range 4-10, while accuracy is in the range 95.7-98.0%. When the second data set was used comparable results were obtained. The observed system performances support the undertaking of system validation in clinical settings.

A neural network expert system for diagnosing and treating hypertension

Hypernet (Hypertension Neural Expert Therapist), a neural network expert system for diagnosing and treating hypertension, is described. After a brief look at artificial neural networks, the authors describe the structure of the three modules that make up Hypernet, starting with the specific problem each network is intended to solve and explaining how the network is expected to operate. The tools developed for implementing the system are a compiler for a simple descriptive language that enables the authors to define, train, and test networks; a graphic editor that translates the network drawn by the user into the proper statements; and a set of programs for interactively generating the examples. The data used as examples, the learning phase, and the results of tests for evaluating the performance of each network are described. and conclusions about overall system results are presented.<<ETX>>

An artificial vision system for X-ray images of human coronary trees

The coronary tree expert (CORTEX) analyzer, which is a vision system for the description of the bidimensional shape and position of coronary vessels using standard nonsubtracted radiographic images, is described. A bottom-up approach was used to deal with the typical characteristics of medical images, such as structural and nonstructural noise and complexity and variability of biological shapes. On these grounds, grouping criteria were utilized to produce intermediate image representations with an increasing complexity in a hierarchical manner (from edge points to curves, segments, bars, and finally to vessels and their mutual relations). In this way, uncertain, inconsistent, and deficient information was efficiently processed. The evaluation of CORTEX segmentation is also performed according to a signal-detection-theory-like approach. >

Tissue characterization from X-ray images

The study of the fine-scale structure of biological tissues is crucial for diagnosing a wide number of different diseases. In X-ray images, fine structures usually induce a correlation among image gray levels and are commonly perceived as textures. In this paper, we report on a Computer Vision approach to the characterization of biological tissues as imaged by standard X-ray techniques. In particular, using features derived from co-occurrence matrices, we have assessed spatial gray-level dependence of bone tissue and lung parenchyma images. A hybrid neural network was adopted to distinguish pathological tissues from normal ones and to classify different pathologies.

Neural network segmentation of magnetic resonance spin echo images of the brain

This paper describes a neural network system to segment magnetic resonance (MR) spin echo images of the brain. Our approach relies on the analysis of MR signal decay and on anatomical knowledge; the system processes two early echoes of a standard multislice sequence. Three main subsystems can be distinguished. The first implements a model of MR signal decay; it synthesizes a four-echo multiecho sequence, in order to add images characterized by long echo-times to the input sequence. The second subsystem exploits a priori anatomical knowledge by producing an image, in which pixels belonging to brain parenchyma are highlighted. Such anatomical information allows the following submodule to distinguish biologically different tissues with similar water content, and hence similar appearance, which might produce misclassifications. The grey levels of the reconstructed sequence and the output of the second module are processed by the third subsystem, which performs the segmentation of the sequence. Each pixel is assigned to one of five different tissue classes that can be revealed with brain MR spin echo imaging. With a suitable encoding, a five-level segmented image can then be produced. The system is based on feed-forward networks trained with the back-propagation algorithm; experiments to assess its performance have been carried out on both simulated and clinical images.

The

Nodule growth as observed in computed tomography (CT) scans acquired at different times is the primary feature to malignancy of indeterminate small lung nodules. In this paper, we propose the estimation of nodule size through a scale-space representation which needs no segmentation and has high intra- and inter-operator reproducibility. Lung nodules usually appear in CT images as blob-like patterns and can be analyzed in the scale-space by Laplacian of Gaussian (LoG ) kernels. For each nodular pattern the LoG scale-space signature was computed and the related characteristic scale adopted as measurement of nodule size. Both in vitro and in vivo validation of LoG characteristic scale were carried out. In vitro validation was done by 40 nondeformable phantoms and 10 deformable phantoms. A close relationship between the characteristic scale and the equivalent diameter, i.e., the diameter of the sphere having the same volume of nodules, (Pearson correlation coefficient was 0.99) and, for nodules undergoing little deformations (obtained at constant volume), small variability of the characteristic scale was observed. The in vivo validation was performed on low and standard-dose CT scans collected from the ITALUNG screening trial (86 nodules) and from the LIDC public data set (89 solid nodules and 40 part-solid nodules or ground-glass opacities). The Pearson correlation coefficient between characteristic scale and equivalent diameter was 0.83-0.93 for ITALUNG and 0.68-0.83 for LIDC data set. Intra- and inter-operator reproducibility of characteristic scale was excellent: on a set of 40 lung nodules of ITALUNG data, two radiologists produced identical results in repeated measurements. The scan-rescan variability of the characteristic scale was also investigated on 86 two-year-stable solid lung nodules (each one observed, on average, in four CT scans) identified in the ITALUNG screening trial: a coefficient of repeatability of about 0.9 mm was observed. Experimental evidence supports the clinical use of the LoG characteristic scale to measure nodule size in CT imaging.

{LoG}

Magnetic resonance and computed tomography produce sets of tomograms which are termed discrete 3D scenes. Usually, discrete 3D scenes are analyzed in two dimensions by observing each tomogram on a screen so that the three-dimensional information contained in the scene can be recovered only partially and qualitatively. The three-dimensional reconstruction of the shape of biological structures from discrete 3D scenes would allow a complete and quantitative recovery of the available information, but this task has proved hard for conventional processing techniques. In this paper we present a system architecture based on neural networks for the fully automated segmentation and recognition of structures of interest in discrete 3D scenes. The system includes a retina and two main processing modules, an Attention-Focuser System and a Region-Finder System, which have been implemented by using feed-forward nets trained with the back-propagation algorithm. This architecture has been tested on computer-simulated structures and has been applied to the reconstruction of the spinal cord and the brain from sets of tomograms.

Characteristic Scale: A Consistent Measurement of Lung Nodule Size in CT Imaging

A general approach to the problem of image matching which exploits a multi-scale representation of local image structure and the principles of self-organizing neural networks is introduced. The problem considered is relevant in many imaging applications and has been largely investigated in medical imagery, especially as regards the integration of different imaging procedures.A given pair of images to be matched, named target and stimulus respectively, are represented by Gabor Wavelets. Correspondence is computed by exploiting the learning procedure of a neural network derived from Kohonens SOM. The SOM units coincide with the pixels of the target image and their weight are pointers to those of the stimulus images. The standard SOM rule is modified so as to account for image features. The properties of our method are tested by experiments performed on synthetic images. The considered implementation has shown that is able to recover a wide range of transformations including global affine transformations and local distortions. Tests in the presence of additive noise indicate considerable robustness against statistical variability. Applications to clinical images are presented.

A neural network architecture for understanding discrete three-dimensional scenes in medical imaging

We report on the use of the entire coronary tree as a natural descriptor of heart motion. The displacement field of the vascular net is estimated by tracking its geometrical and topological features. Afterwards, the shape and Me movement of epicardium are recovered by using a deformable model derived on the basis of general physical assumptions. The heart surface and the related motion field are represented through a spherical harmonic series. This permits a compact analytical description which is global and focal at the same time and is well suited for a multi-level analysis. Experimental results on clinical angiographic data are illustrated.

Matching of medical images by self-organizing neural networks

The authors describe a fully automated neural-network-based computer-vision system for recovering the endocardial left ventricular surface from a reduced number of echocardiograms. After detecting and linking edges, an edge-segment representation is built for each input image. According to their parameters, edge-segment primitives are then labeled as noise or LV-boundary by a feedforward neural net. The LV-boundary segments are utilized in defining the energy function of a continuous Hopfield neural network. This function is a finite-element approximation of the potential energy of a closed elastic thin-surface in the presence of external forces. The network relaxes into a minimum energy state which represents the reconstructed LV surface. Experimental results are reported which show how this surface description can be used for displaying the static and dynamic LV shape and for measuring parameters such as LV volume.<<ETX>>