Automated detection of mass lesions in dedicated breast CT: a preliminary study.

PURPOSE: To develop an automated method to detect breast masses on dedicated breast CT (BCT) volumes and to conduct a preliminary evaluation of its performance. This method can be used in a computer-aided detection (CADe) system for noncontrast enhanced BCT. METHODS: The database included patient images, which were acquired under an IRB-approved protocol. The database in this study consisted of 132 cases. 50 cases contained 58 malignant masses, and 23 cases contained 24 benign masses. 59 cases did not contain any biopsy-proven lesions. Each case consisted of an unenhanced CT volume of a single breast. First, each breast was segmented into adipose and glandular tissues using a fuzzy c-means clustering algorithm. The glandular breast regions were then sampled at a resolution of 2 mm. At each sampling step, a 3.5-cm(3) volume-of-interest was subjected to constrained region segmentation and 17 characteristic features were extracted, yielding 17 corresponding feature volumes. Four features were selected using step-wise feature selection and merged with linear discriminant analysis trained in the task of distinguishing between normal breast glandular regions and masses. Detection performance was measured using free-response receiver operating characteristic analysis (FROC) with leave-one-case-out evaluation. RESULTS: The feature selection stage selected features that characterized the shape and margin strength of the segmented region. CADe sensitivity per case was 84% (std = 4.2%) at 2.6 (std = 0.06) false positives per volume, or 6 × 10(-3) per slice (at an average of 424 slices per volume in this data set). CONCLUSIONS: This preliminary study demonstrates the feasibility of our approach for CADe for BCT.

Computerized mass detection for digital breast tomosynthesis directly from the projection images.

Digital breast tomosynthesis (DBT) has recently emerged as a new and promising three-dimensional modality in breast imaging. In DBT, the breast volume is reconstructed from 11 projection images, taken at source angles equally spaced over an arc of 50 degrees. Reconstruction algorithms for this modality are not fully optimized yet. Because computerized lesion detection in the reconstructed breast volume will be affected by the reconstruction technique, we are developing a novel mass detection algorithm that operates instead on the set of raw projection images. Mass detection is done in three stages. First, lesion candidates are obtained for each projection image separately, using a mass detection algorithm that was initially developed for screen-film mammography. Second, the locations of a lesion candidate are backprojected into the breast volume. In this feature volume, voxel intensities are a combined measure of detection frequency (e.g., the number of projections in which a given lesion candidate was detected), and a measure of the angular range over which a given lesion was detected. Third, features are extracted after reprojecting the three-dimensional (3-D) locations of lesion candidates into projection images. Features are combined using linear discriminant analysis. The database used to test the algorithm consisted of 21 mass cases (13 malignant, 8 benign) and 15 cases without mass lesions. Based on this database, the algorithm yielded a sensitivity of 90% at 1.5 false positives per breast volume. Algorithm performance is positively biased because this dataset was used for development, training, and testing, and because the number of algorithm parameters was approximately the same as the number.of patient cases. Our results indicate that computerized mass detection in the sequence of projection images for DBT may be effective despite the higher noise level in those images.

Computerized characterization of mammographic masses: analysis of spiculation.

Although general rules for the differentiation between benign and malignant breast lesions exist, only 10 to 20% of masses referred for surgical breast biopsy are actually malignant. We are developing, as an aid to radiologists, a computerized scheme for the classification of masses appearing on mammograms to reduce the number of false-positive diagnoses of malignancies. The classification scheme involves the extraction of the margin of masses in order to quantify the degree of spiculation, which, in turn, is related to the likelihood of malignancy. When two measures of spiculation are used as input to an artificial neural network, the scheme achieves a performance similar to that achieved when radiologist's spiculation ratings alone are used for a clinical database of 53 masses. The computerized classification scheme therefore has the potential to effectively aid radiologists in determining appropriate patient management.

Computerized detection of pulmonary nodules in computed tomography images.

RATIONALE AND OBJECTIVES: Interpretation of computed tomographic (CT) scans of the lungs is a time-consuming task that involves visual correlation of possible nodules in one section with those in contiguous sections to distinguish actual nodules from blood vessels. Thus, the authors are developing automated methods to detect nodules on CT images of the thorax. METHODS: The computerized technique uses various computer-vision techniques and a priori information of the morphologic characteristics of pulmonary nodules. In each section, the external thoracic wall and lung boundaries are detected, and the features within the lung boundaries are subjected to gray-level thresholding operations. By analyzing the relationships between features arising at different threshold levels with respect to their shape, size, and location, each feature is assigned a likelihood of being a nodule or a vessel. Features in adjacent sections are compared to resolve ambiguous features. Detected nodule candidates are displayed in three dimensions within the lung. RESULTS: The system provided a sensitivity of 94% for nodule detection and an average of 1.25 false-positive results per case. CONCLUSIONS: Continued development of an automated method for detecting pulmonary nodules in CT scans is expected to aid radiologists in the task of locating nodules in three dimensions.

Comparison of bilateral-subtraction and single-image processing techniques in the computerized detection of mammographic masses.

RATIONALE AND OBJECTIVES: Identification of regions as possible masses on digitized screen film mammograms is an important initial step in the computerized detection of breast carcinomas. Possible masses may be initially extracted using criteria based on optical densities, geometric patterns, and asymmetries between corresponding locations in right and left mammograms. In this study, the usefulness of information arising from mammographic asymmetries for the identification of mass lesions is investigated. METHODS: Two techniques are investigated--a nonlinear bilateral-subtraction technique based on image pairs and a local gray-level thresholding technique based on single images. Detection performances obtained with the two techniques in combination with various feature-analysis techniques are evaluated using 154 pairs of mammograms and compared using free-response receiver operating characteristic (FROC) analysis. RESULTS: The nonlinear bilateral-subtraction technique performed better than the local gray-level thresholding technique. CONCLUSION: The incorporation of asymmetric information appears to be useful for computerized identification of possible masses on mammograms.

Computerized scheme for the detection of pulmonary nodules. A nonlinear filtering technique.

To aid radiologists in the detection of lung cancer, the authors are developing a computer-aided diagnosis system that locates areas suspicious for nodules in digital chest radiographs. The system involves a difference-image approach and various feature-extraction techniques. The authors describe nonlinear filters used in the difference-image approach. A morphological open operation and a ring-shaped median filter are applied in the difference-image step for signal enhancement and signal suppression, respectively. Using 60 clinical chest radiographs, the nonlinear filtering method detected approximately 63% of actual nodules with approximately 19 false-positive results per image. The locations of the false-positive detections, however, usually did not coincide with those from the linear filtering method. Thus, by using a combination of the detections from the two methods, the false-positive rate was reduced to two to three per image at a sensitivity of 60%.

Potential usefulness of computerized nodule detection in screening programs for lung cancer.

RATIONALE AND OBJECTIVE: To alert radiologists to possible nodule locations and subsequently to reduce the number of false-negative diagnoses, the authors are developing a computer-aided diagnostic (CAD) scheme for the detection of lung nodules in digital chest images. METHODS: A computer-vision scheme was applied to photofluorographic films obtained in a mass survey for detection of asymptomatic lung cancer in Japan. Ninety-five patients with abnormal test results who had primary and metastatic lung cancers and 103 patients with normal test results were included. RESULTS: The sensitivity of the computer output was comparable with that of physicians in this mass survey (62%). The computer detected approximately 40% of all nodules missed in the mass survey, but missed 17 true-positive results identified in the mass survey. The CAD scheme produced an average of 15 false-positive findings per image. CONCLUSION: If the number of false-positive results can be significantly reduced, computer-vision schemes such as this may have a role in lung cancer screening programs.

Computer-aided diagnosis in chest radiography. Preliminary experience.

RATIONALE AND OBJECTIVES: Computer-aided diagnosis (CAD) schemes for chest radiography are being developed with which to alert radiologists to possible lesions, and thus potentially improve diagnostic accuracy. However, CAD schemes have not been tested on a large number of clinical cases. The authors identify design parameters that would be required for development of an intelligent workstation. METHODS: Computer-aided diagnosis programs were applied for the automated detection of lung nodules, cardiomegaly, and interstitial infiltrates to 310 consecutive chest radiographs, and were analyzed for potential usefulness and limitations. Computer-aided diagnosis output was evaluated by radiologists and physicists for accuracy and technical problems, respectively. RESULTS: Approximately 70% of the results were judged to be potentially acceptable; however, the number of false-positive findings was relatively high. Technical problems included failure to detect subtle abnormalities and the occurrence of false-positive detections caused by normal anatomical structures. CONCLUSION: Computer-aided diagnosis has the potential to be a valuable aid to radiologists in clinical practice, if certain technical problems can be overcome and if optimal operating points can be defined for clinical use.

Image feature analysis of false-positive diagnoses produced by automated detection of lung nodules.

RATIONALE AND OBJECTIVES: To reduce the number of false-negative diagnoses by radiologists, the authors are developing a computer-aided diagnosis scheme for detection of lung nodules in digital chest images. In this study, the authors attempted to reduce the number of false-positive diagnoses obtained with a previous computer scheme by incorporating additional knowledge from experienced chest radiologists into the computer scheme. METHODS: The authors applied their previous computer scheme, using less-strict criteria, to 60 clinical chest radiographs; this yielded 735 candidate nodules (23 true nodules and 712 false-positive diagnoses). These candidates were analyzed using region-growing, trend-correction, and edge-gradient techniques to determine measures by which to quantify image features of candidate nodules. RESULTS: The 712 false-positive diagnoses represented various anatomic structures that were located throughout the chest image. From this analysis, we were able to decrease the number of false-positive errors from an average of 12 to approximately 5 per image without eliminating any true nodules. CONCLUSION: Our results show that incorporating knowledge from experienced chest radiologists into the computer algorithm will play an important role in the development of computerized schemes for the detection of pulmonary nodules.

Basic imaging properties of a large image intensifier-TV digital chest radiographic system.

The basic imaging properties of a large (57 cm) image intensifier (I.I.)-TV digital imaging system were examined to determine the effects of various physical parameters on the quality of the digital chest images obtained, and also to explore the clinical usefulness of the system. The characteristic curve of the digital system, which relates the output pixel value to the input relative x-ray intensity, was measured with an aluminum stepwedge. MTFs were determined using slit images, and the veiling-glare fraction was measured with a lead-disk technique. Noise Wiener spectra were obtained from uniformly exposed images. The current limitations of the large II-TV digital chest system are its low spatial resolution, and the presence of large amounts of veiling glare and structure mottle. Advantages of this system over other digital chest imaging systems include the high speed of image data acquisition and the capability of "real-time" dynamic imaging of the chest at a radiation dose comparable to that in conventional radiography of the chest.

Feature selection with limited datasets.

Computer-aided diagnosis has the potential of increasing diagnostic accuracy by providing a second reading to radiologists. In many computerized schemes, numerous features can be extracted to describe suspect image regions. A subset of these features is then employed in a data classifier to determine whether the suspect region is abnormal or normal. Different subsets of features will, in general, result in different classification performances. A feature selection method is often used to determine an "optimal" subset of features to use with a particular classifier. A classifier performance measure (such as the area under the receiver operating characteristic curve) must be incorporated into this feature selection process. With limited datasets, however, there is a distribution in the classifier performance measure for a given classifier and subset of features. In this paper, we investigate the variation in the selected subset of "optimal" features as compared with the true optimal subset of features caused by this distribution of classifier performance. We consider examples in which the probability that the optimal subset of features is selected can be analytically computed. We show the dependence of this probability on the dataset sample size, the total number of features from which to select, the number of features selected, and the performance of the true optimal subset. Once a subset of features has been selected, the parameters of the data classifier must be determined. We show that, with limited datasets and/or a large number of features from which to choose, bias is introduced if the classifier parameters are determined using the same data that were employed to select the "optimal" subset of features.

Investigation of basic imaging properties in digital radiography. 3. Effect of pixel size on SNR and threshold contrast.

The effect of pixel size on the signal-to-noise ratio (SNR) and threshold detection of low-contrast radiologic patterns was investigated theoretically for digital radiographic systems. The SNR based on the perceived statistical decision theory model, together with the internal noise of the human eye-brain system, was calculated by using two-dimensional displayed digital signal spectra and noise Wiener spectra. Threshold contrasts were predicted from the calculated SNR for various combinations of object size and shape, pixel size, resolution, and noise. Predicted threshold contrasts agreed well with those determined experimentally in an observer performance study. The threshold contrast of small objects increased substantially as the pixel size increased beyond 0.2 mm. For pixel sizes of 0.1 and 0.2 mm, however, the threshold contrasts were similar. Since a digital system is not shift invariant, a range of threshold contrast results for a small object and a large pixel, depending on the alignment of the object position relative to the sampling coordinates.

Investigation of basic imaging properties in digital radiography. I. Modulation transfer function.

The effect of various digital parameters, such as the sampling aperture, sampling distance, and display aperture, on the modulation transfer function (MTF) of digital radiographic imaging systems was investigated by means of theoretical simulation studies. The MTFs were also determined experimentally to confirm the relationship used in the simulation studies. The results indicate that the overall MTF of a digital system cannot specify the resolution properties in the same way as can the MTFs of analog systems. The MTF of a digital system may include a "false" response due to aliasing, which could lead to an incorrect interpretation of the resolution properties. The magnitude of aliasing that will occur in a digitized signal depends on the sampling parameters chosen and on the frequency content of the radiologic object being imaged. Thus, the type of object to be detected as well as various digital parameters must be considered in the design and evaluation of digital imaging systems.

Computerized detection of abnormal asymmetry in digital chest radiographs.

A technique for automated detection of abnormal asymmetry in digital chest radiographs is being developed. Such a method could be used to prescreen chest radiographs to bring obviously abnormal cases to the immediate attention of a radiologist. In addition, this technique may be used to detect large-area abnormalities which may cause other, more lesion-specific computer algorithms to fail. Asymmetric abnormalities are detected by multiple stages of gray-level thresholding. Lung contours are determined, and after a centroid test is used to eliminate contours external to the lungs, the areas of remaining contours are calculated. The present scheme, applied to a database of 70 chest images, yielded a sensitivity of 91% and a specificity of 80%.

Multifractal radiographic analysis of osteoporosis.

An important complication of osteoporosis is fracture. Alteration in bone structure, as well as decreased bone mass, contribute to the tendency to fracture in osteoporosis. Current methods that measure bone mass alone show substantial overlap of the measurements of osteoporotic patients who fracture with those that do not. Our aim is to develop noninvasive methods of evaluating bone structure on plain film radiographs to better predict fracture risk in osteoporosis. Regions of interest (ROIs) were selected from digitized lateral lumbar spine radiographs of 43 patients being seen in an osteoporosis clinic. The fractal dimension of these ROIs was estimated using a surface area method. The ability of fractal dimension to distinguish between cases that had fracture elsewhere in the spine from those that did not, was evaluated using receiver operating characteristic (ROC) analysis. These results were compared with ROC analysis for these same patients using bone mineral density (BMD) measurements (bone mass). Significantly larger Az (area under ROC curve) values were obtained using fractal dimension (0.87) than from using BMD (0.58), indicating a better test performance using fractal dimension. Therefore, computerized radiographic methods to evaluate bone structure, such as fractal analysis, may be helpful in better determining fracture risk in osteoporosis.

Investigation of basic imaging properties in digital radiography. 7. Noise Wiener spectra of II-TV digital imaging systems.

We used Wiener spectral analysis in order to investigate the different noise sources and the effects of various parameters such as pixel size, image intensifier (II) field size, and exposure level on the noise in an II-TV digital system. The digital Wiener spectra in terms of relative x-ray intensity were determined directly from the digital noise data in terms of pixel values, by use of the characteristic curve of the imaging system. From averaged, subtracted, and/or combination images, the amount of structure mottle relative to the amount of quantum mottle was estimated. We found that a substantial amount of structure mottle was included in our II-TV digital subtraction angiography system, whereas the electronic noise of the TV system was quite small relative to the quantum and structure mottle. The effects of time jitter on the noise in single-frame images (consisting of multiple video frames) and in subtracted and averaged images were also investigated.

Investigation of basic imaging properties in digital radiography. 9. Effect of displayed grey levels on signal detection.

Results of an 18-alternative forced-choice experiment have shown that observers were capable of detecting a signal with a contrast of 1 in terms of 10-bit data which were displayed on a CRT monitor with an 8-bit video generator and a window width setting of 1024. We investigated the conditions under which 10-bit signal data can be detected when displayed using an 8-bit video generator. Results show that the 10-bit digital quantum noise, which was approximately Gaussian distributed, can act as a carrier of the signal data, thus allowing a signal having a contrast of a fraction of a displayed grey level to be detected. We demonstrate the relationship between the rms value of the digital noise (obtainable with a clinical digital subtraction angiography system under various exposure levels), the number of bits available in the display video generator, and the "transmitted" signal contrast displayed on the CRT monitor.

Automated detection of lung nodules in CT scans: preliminary results.

We have developed a fully automated computerized method for the detection of lung nodules in helical computed tomography (CT) scans of the thorax. This method is based on two-dimensional and three-dimensional analyses of the image data acquired during diagnostic CT scans. Lung segmentation proceeds on a section-by-section basis to construct a segmented lung volume within which further analysis is performed. Multiple gray-level thresholds are applied to the segmented lung volume to create a series of thresholded lung volumes. An 18-point connectivity scheme is used to identify contiguous three-dimensional structures within each thresholded lung volume, and those structures that satisfy a volume criterion are selected as initial lung nodule candidates. Morphological and gray-level features are computed for each nodule candidate. After a rule-based approach is applied to greatly reduce the number of nodule candidates that corresponds to nonnodules, the features of remaining candidates are merged through linear discriminant analysis. The automated method was applied to a database of 43 diagnostic thoracic CT scans. Receiver operating characteristic (ROC) analysis was used to evaluate the ability of the classifier to differentiate nodule candidates that correspond to actual nodules from false-positive candidates. The area under the ROC curve for this categorization task attained a value of 0.90 during leave-one-out-by-case evaluation. The automated method yielded an overall nodule detection sensitivity of 70% with an average of 1.5 false-positive detections per section when applied to the complete 43-case database. A corresponding nodule detection sensitivity of 89% with an average of 1.3 false-positive detections per section was achieved with a subset of 20 cases that contained only one or two nodules per case.

Investigation of basic imaging properties in digital radiography. 6. MTFs of II-TV digital imaging systems.

We devised a new, simple technique for measuring the modulation transfer function (MTF) of a digital imaging system by using an image of an angulated slit. With this technique, the "presampling" analog MTF, which includes the geometric unsharpness, the detector unsharpness, and the unsharpness of the sampling aperture, can be measured even beyond the Nyquist frequency. A single-frame image of a slightly angulated slit was employed in order to obtain Fourier transforms of line spread functions at different alignments. The presampling MTF was determined by averaging the two Fourier transforms which we obtained from two extreme alignments (center and shifted) of the slit relative to the sampling coordinate. The presampling MTFs of our digital subtraction angiographic system were determined in two orthogonal directions for three different image-intensifier modes.

Computerized analysis of breast lesions in three dimensions using dynamic magnetic-resonance imaging.

Contrast-enhanced magnetic resonance imaging (MRI) of the breast is known to reveal breast cancer with higher sensitivity than mammography alone. The specificity is, however, compromised by the observation that several benign masses take up contrast agent in addition to malignant lesions. The aim of this study is to increase the objectivity of breast cancer diagnosis in contrast-enhanced MRI by developing automated methods for computer-aided diagnosis. Our database consists of 27 MR studies from 27 patients. In each study, at least four MR series of both breasts are obtained using FLASH three-dimensional (3D) acquisition at 90 s time intervals after injection of Gadopentetate dimeglumine (Gd-DTPA) contrast agent. Each series consists of 64 coronal slices with a typical thickness of 2 mm, and a pixel size of 1.25 mm. The study contains 13 benign and 15 malignant lesions from which features are automatically extracted in 3D. These features include margin descriptors and radial gradient analysis as a function of time and space. Stepwise multiple regression is employed to obtain an effective subset of combined features. A final estimate of likelihood of malignancy is determined by linear discriminant analysis, and the performance of classification by round-robin testing and receiver operating characteristics (ROC) analysis. To assess the efficacy of 3D analysis, the study is repeated in two-dimensions (2D) using a representative slice through the middle of the lesion. In 2D and in 3D, radial gradient analysis and analysis of margin sharpness were found to be an effective combination to distinguish between benign and malignant masses (resulting area under the ROC curve: 0.96). Feature analysis in 3D was found to result in higher performance of lesion characterization than 2D feature analysis for the majority of single and combined features. In conclusion, automated feature extraction and classification has the potential to complement the interpretation of radiologists in an objective, consistent, and accurate way.