Jyh-Shyan Lin

Reduction of false positives in lung nodule detection using a two-level neural classification

The authors have developed a neural-digital computer-aided diagnosis system, based on a parameterized two-level convolution neural network (CNN) architecture and on a special multilabel output encoding procedure. The developed architecture was trained, tested, and evaluated specifically on the problem of diagnosis of lung cancer nodules found on digitized chest radiographs. The system performs automatic suspect localization, feature extraction, and diagnosis of a particular pattern-class aimed at a high degree of true-positive fraction detection and low false-positive fraction detection. In this paper, the authors aim at the presentation of the two-level neural classification method in reducing false-positives in their system. They employed receiver operating characteristics (ROC) method with the area under the ROC curve (A(z)) as the performance index to evaluate all the simulation results. The two-level CNN showed superior performance (A(z)=0.93) to the single-level CNN (A(z)=0.85). The proposed two-level CNN architecture is proven to be promising and to be extensible, problem-independent, and therefore, applicable to other medical or difficult diagnostic tasks in two-dimensional (2-D) image environments.

Digital mammography: tradeoffs between 50- and 100-micron pixel size

Improvements in mammography equipment related to a decrease in pixel size of digital mammography detectors raise questions of the possible effects of these new detectors. Mathematical modeling suggested that the benefits of moving from 100 to 50 micron detectors were slight and might not justify the cost of these new units. Experiments comparing screen film mammography, a storage phosphor 100 micron digital detector, a 50 micron digital breast spot device, 100 micron film digitization and 50 micron film digitization suggests that object conspicuity should be better for digital compared to conventional systems, but that there seemed to be minimal advantage to going from 100 to 50 microns. The 50 micron pixel system appears to provide a slight advantage in object contrast and perhaps in shape definition, but did not allow smaller objects to be detected.

Application of artificial neural networks for reducing false positives in lung nodule detection on digital chest radiographs

The objective of many existing computer-aided diagnosis (CADx) schemes for lung nodule detection is to reduce the number of false-positives (i.e., increase specificity) while maintaining a high level of sensitivity. Our examination of the false-positives obtained with the previously developed CADx program show that many round objects, such as rib crossings, end-on vessels, and aggregates of vessels, were mistakenly classified as nodules. Among the problems of decreasing the number of false-positives, the differentiation between nodules and end-on vessels is one of the most challenging tasks performed by computers. To eliminate the false-positives, two methods are proposed. One method is to extract the known features (i.e., contrast and size) based on a conventional digital image processing technique. The other method uses an artificial neural network (ANN) which is specifically trained to classify nodules and end-on vessels. Performances of the two approaches are evaluated using the receiver operating characteristics (ROC) method and the area under the ROC curve (Az). Based on our test database, the FFNN and the algorithmic approaches showed preliminary ROC performances with Az values equal to 0.90 and 0.94, respectively.

Detection of clustered microcalcifications using fuzzy modeling and convolution neural network

This paper describes an automatic computer searching system for detecting clustered microcalcifications. A fuzzy classification modeling was employed to extract each suspected microcalcification possessing similar physical parameters. Therefore, only those possible classes were evaluated using a sophisticated convolution neural network which requires a great deal of computation and serves as a discriminator. Based on the detected spots, many of them are true microcalcifications, the computer can easily make a determination when 5 spots are located within a defined region. However, when a cluster consists of only two to four suspicious spots a fuzzy function was used to determine the inclusion of other spots near the cluster. This can be very important for the detection of subtle cases. The membership of the latter fuzzy function was composed of the distance between the suspected spots as well as the output values of the convolution neural network. We have tested the improved algorithms on our research database consisting of 45 mammograms. The results indicated that the fuzzy classification modeling decreased the number of false-positives from 2,874 to 1,067 suspected spots per image without increasing any false-negative detection. The over-all performance in the detection of clustered microcalcifications through the updated algorithms was 90% sensitivity at 0.5 false-positive per image. The computation time using a DEC-Alpha workstation was decreased from 5 minutes to about 3 minutes per image.

Computer-assisted diagnosis for lung nodule detection using a neural network technique

The potential advantages of using digital techniques instead of film-based radiology have been discussed very extensively for the past ten years. These advantages are found mainly in the computer management of picture archiving and communication systems (PACS). On the other hand, the computer-assisted diagnosis (CADx) could potentially enhance radiological services in the future. Lung nodule detection has been a clinically difficult subject for many years. Most of the literature has indicated that the finding rate for lung nodules (size range from 3 mm to 15 mm) is only about 65%, and 30% of the missing nodule can be found retrospectively. In the recent research, imaging processing techniques, such as thresholding and morphological analysis, have been employed to enhance the true-positive detection. However, these methods still produce many false-positive detections. We have used neural networks to distinguish true-positives from the suspected areas-of-interest which are generated from signal enhanced image. The initial results show that the trained neural networks program can increase true-positive detections and drastically reduce the number of false-positive detections. This program can perform three modes of lung nodule detection: (1) thresholding, (2) profile matching analysis, and (3) neural network. This program is fully automatic and has been implemented in a DEC 5000/200 workstation. The total processing time for all three methods is less than 35 seconds. We are planning to link this workstation to our PACS for further clinical evaluation. In this paper, we report our neural network and fast algorithms for various image processing techniques for the lung nodule detection and show the results of the initial studies.

Region-based enhancement of chest and cervical spine radiographs

We have developed a region-based image processing method to enhance selective radiodense regions on digital radiographs. We employ a wavelet filtering technique to locate the radiodense regions-of-interest and then apply different degrees of enhancement procedure to them. The enhancement procedure is based on an unsharp masking technique controlled by a set of sigmoidal functions. The method was tested on computed chest radiographs to improve the visualization of the mediastinum and radiodense spine areas. The enhanced chest images showed improved visualization in the mediastinum area, and the visibility of vascular structures which were obscured by the diaphragm and mediastinum was improved. To demonstrate the methods potential in other medical image processing tasks, we applied it to cervical spine images. The processed cervical spine images also showed better visualization of the seventh cervical vertebrae and the first thoracic vertebrae in the high radiodense area caused by the superimposition of the patients shoulder tissue over these regions of interest.

Digital mammography: the effects of decreased exposure

It has been stated that digital mammography will reduce the exposure required for mammography. This poster explores the effects of decreased exposure on the information present in digital mammography. In general, the digital system performed better than screen film mammography with lower exposures. With the usual exposures used for screen film mammography, performance was equal. With high exposures sufficient to result in a dark film (OD 1.5), the digital system performed better than screen film with very small test objects. Proposals have been made to decrease the tube loading required for slot scanning devices by increasing KVP. This would result in their being less object contrast due to the decreases in the absorption coefficient of calcium compared to water at higher KVP. This poster looks at the potential for correcting the loss in object contrast that would result from the use of high contrast look up tables. It was found that in the tested system, one could restore the information in one of the two test objects used (but not the other), but that the image processing methods used would result in an image that radiologists would probably find inadequate for interpretation.

Digital mammography in the radio-dense and complex pattern breast

The sensitivity of mammography for the detection of breast cancer is decreased in the radiodense breast. Storage phosphor digital radiographic systems have a wider latitude of exposure than conventional mammographic screen film systems. By using low resolution histogram equalization one can produce a mammographic image of the breast that retains the high frequency information that defines the edges of microcalcifications, architectural distortion and some masses but which, at the same time, allows one to look through into regions of increased breast radiodensity and identify microcalcifications within them. This paper demonstrates the effect of this special form of image processing.

Design-based approach to ethics in computer-aided diagnosis

A design-based approach to ethical analysis examines how computer scientists, physicians and patients make and justify choices in designing, using and reacting to computer-aided diagnosis (CADx) systems. The basic hypothesis of this research is that values are embedded in CADx systems during all phases of their development, not just retrospectively imposed on them. This paper concentrates on the work of computer scientists and physicians as they attempt to resolve central technical questions in designing clinically functional CADx systems for lung cancer and breast cancer diagnosis. The work of Lo, Chan, Freedman, Lin, Wu and their colleagues provides the initial data on which this study is based. As these researchers seek to increase the rate of true positive classifications of detected abnormalities in chest radiographs and mammograms, they explore dimensions of the fundamental ethical principal of beneficence. The training of CADx systems demonstrates the key ethical dilemmas inherent in their current design.

Artificial convolution neural network with wavelet kernels for disease pattern recognition

A two-dimensional convolution neural network (CNN) with wavelet kernels (WK) has been developed for image pattern recognition. The structure of the CNN is a simplified version of the neocognitron. We used only a two-level structure and eliminated all complex-cell layers. Nets between two adjacent layers in the feature selection level of the CNN are selectively interconnected across groups. In this part of the CNN signals processing, each group in the receiving layer receives signals from a group of weights (i.e., kernels). For the forward signal propagation, the product obtained from the kernel convoluting the front layer is collected onto the corresponding matrix element of the receiving layer. In this paper, the convolution kernels of the CNN (CNN/WK) are wavelet based and are trained by a supervised manner. In the development of the CNN/WK, we forced each updated convolution kernel to be orthonormal. Therefore, features (transformed coefficients) selected on the transform domain are linearly independent. Hence, the fully connected layers in the classification level of the CNN can perform more effectively. The applications of the CNN for disease pattern recognition have been very successful. When isolated patterns were further processed by internal filtering and classification layers were built into the neural network structure, the disease patterns were more easily recognized. Although, we did not receive substantial improvement of the ROC performance using the CNN/WK, this method may assist us in the analysis of the trained kernels and eventually lead to the optimization of feature extraction in a course of disease pattern recognition.