Computerized analysis of mammographic microcalcifications in morphological and texture feature spaces

We are developing computerized feature extraction and classification methods to analyze malignant and benign microcalcifications on digitized mammograms. Morphological features that described the size, contrast, and shape of microcalcifications and their variations within a cluster were designed to characterize microcalcifications segmented from the mammographic background. Texture features were derived from the spatial gray-level dependence (SGLD) matrices constructed at multiple distances and directions from tissue regions containing microcalcifications. A genetic algorithm (GA) based feature selection technique was used to select the best feature subset from the multi-dimensional feature spaces. The GA-based method was compared to the commonly used feature selection method based on the stepwise linear discriminant analysis (LDA) procedure. Linear discriminant classifiers using the selected features as input predictor variables were formulated for the classification task. The discriminant scores output from the classifiers were analyzed by receiver operating characteristic (ROC) methodology and the classification accuracy was quantified by the area, Az, under the ROC curve. We analyzed a data set of 145 mammographic microcalcification clusters in this study. It was found that the feature subsets selected by the GA-based method are comparable to or slightly better than those selected by the stepwise LDA method. The texture features (Az = 0.84) were more effective than morphological features (Az = 0.79) in distinguishing malignant and benign microcalcifications. The highest classification accuracy (Az = 0.89) was obtained in the combined texture and morphological feature space. The improvement was statistically significant in comparison to classification in either the morphological (p = 0.002) or the texture (p = 0.04) feature space alone. The classifier using the best feature subset from the combined feature space and an appropriate decision threshold could correctly identify 35% of the benign clusters without missing a malignant cluster. When the average discriminant score from all views of the same cluster was used for classification, the Az value increased to 0.93 and the classifier could identify 50% of the benign clusters at 100% sensitivity for malignancy. Alternatively, if the minimum discriminant score from all views of the same cluster was used, the Az value would be 0.90 and a specificity of 32% would be obtained at 100% sensitivity. The results of this study indicate the potential of using combined morphological and texture features for computer-aided classification of microcalcifications.     

Perinatal outcome and the social contract--interrelationships between health and humanity

OBJECTIVE: To assess an automated algorithm, developed for the classification of normal and cancerous colonic mucosa, using geometric analysis of features and texture analysis. STUDY DESIGN: Twenty-one images were analyzed, 10 from normal and 11 from cancerous mucosa. The classification was based on a regularity index dependent on shape, object orientation for establishing parallelism and five texture features derived using the co-occurrence image analysis method. RESULTS: Geometric analysis yielded an overall classification accuracy of 80%. The corresponding sensitivity and specificity were 94% and 64%, respectively. Using texture analysis, the overall classification accuracy was 90%, with a sensitivity and specificity of 82% and 100%, respectively. CONCLUSION: This initial study demonstrated that geometric and texture analysis techniques show promise for automated analysis of colon cancer.     

Regarding Kumar et al. IJROBP 31(3):477-483; 1995

OBJECTIVE: We present a computer-aided diagnostic technique for identifying nodular interstitial lung disease on chest radiographs. The fractal dimension was used as a numerical measure of image texture on digital chest radiographs to distinguish patients with normal lung from those with a diffuse nodular interstitial abnormality. MATERIALS AND METHODS: Twenty digitized chest radiographs were classified as normal (n = 10) or as containing diffuse nodular abnormality (n = 10) on the basis of readings assigned according to the classification of the International Labour Organization. Regions of interest (ROIs) measuring 1.28 cm2 were selected from the intercostal spaces of these radiographs. The fractal dimension of these ROIs was estimated by power spectrum analysis. The cases were not subtle. RESULTS: The fractal dimension provided statistically significant discrimination between normal parenchyma and nodular interstitial lung disease. The area under the receiver operating characteristic curve was 0.90 (+/- 0.02). One operating point provides sensitivity of 88% with a specificity of 80%. CONCLUSION: The fractal dimension can provide a measure of lung parenchymal texture and shows promise as an element of computer-aided diagnosis, characterization, and follow-up of interstitial lung disease.     

Wavelet image extension for analysis and classification of infarcted myocardial tissue

Some computer applications for tissue characterization in medicine and biology, such as analysis of the myocardium or cancer recognition, operate with tissue samples taken from very small areas of interest. In order to perform texture characterization in such an application, only a few texture operators can be employed: the operators should be insensitive to noise and image distortion and yet be reliable in order to estimate texture quality from the small number of image points available. In order to describe the quality of infarcted myocardial tissue, we propose a new wavelet-based approach for analysis and classification of texture samples with small dimensions. The main idea of this method is to decompose the given image with a filter bank derived from an orthonormal wavelet basis and to form an image approximation with higher resolution. Texture energy measures calculated at each output of the filter bank as well as energies of synthesized images are used as texture features in a classification procedure. We propose an unsupervised classification technique based on a modified statistical t-test. The method is tested with clinical data, and the classification results obtained are very promising. The performance of the new method is compared with the performance of several other transform-based methods. The new algorithm has advantages in classification of small and noisy input samples, and it represents a step toward structural analysis of weak textures.     

Computer-aided diagnosis for interstitial infiltrates in chest radiographs: optical-density dependence of texture measures

We have been developing a computerized scheme for automated detection and characterization of interstitial infiltrates based on the Fourier transform of lung texture. To improve the performance of the scheme, which was developed using digitized screen-film radiographs, optical-density dependence of both the gradient of the film used and the system noise associated with the laser scanner were investigated. Two hundred chest radiographs, including 100 abnormal cases with interstitial infiltrates, were digitized using a laser scanner. The root-mean-square (RMS) variations and the first moments of the power spectra, which correspond to the magnitude and coarseness of lung texture, were determined by Fourier transform of lung textures in numerous regions of interest (ROIs). The RMS variation was dependent upon the average optical density in the ROI, though no obvious trend existed for the first moment of the power spectrum. Dependence of the RMS variations on optical density was corrected for using the gradient curve of the film. Also, system noise associated with the laser scanner was corrected. Results indicated that the specificity was improved from 81% (without correction) to 89% (with corrections), without any loss of sensitivity (90%). Thus, the correspondence between the computer output and consensus interpretation of radiologists was improved with the new scheme compared to the previous one. This improved computerized scheme may be useful to radiologists in detecting interstitial infiltrates in chest radiographs.     

Quantitative brain SPECT in Alzheimer's disease and normal aging

To improve the diagnostic utility of brain single-photon emission computed tomography (SPECT) in Alzheimer's disease (AD), we have developed and evaluated an objective method of differentiating patients and healthy elderly controls using a quantitative image analysis protocol. HMPAO-SPECT image datasets from 29 patients with probable AD and 78 age-matched controls were registered to a common anatomic frame of reference. Activity levels within 120 standardized cortical volumes were determined by an automated procedure. Subjects were classified into normal and AD groups by quadratic discriminant analysis using two features: global average activity level and average normalized activity levels within the two clusters of standardized volumes identified as most significantly different in AD by analysis of covariance. The classification used split-half replication to ensure valid results. Classification performance quantified by the area under a binormal ROC curve fitted to the data was 0.923 +/- 0.036; at a threshold likelihood ratio of 1, the sample sensitivity was 91% and specificity was 86%. We conclude that quantitative SPECT accurately distinguishes AD patients from elderly controls.     

Computerized detection of masses in digitized mammograms using single-image segmentation and a multilayer topographic feature analysis

RATIONALE AND OBJECTIVES: We developed and evaluated a computer-aided detection (CAD) scheme for masses in digitized mammograms. METHODS: A multistep CAD scheme was developed and tested. The method uses a technique of single-image segmentation with Gaussian bandpass filtering to yield a high sensitivity for mass detection. A rule-based multilayer topographic feature analysis method is then used to classify suspected regions. A set of 260 cases, including 162 verified masses, was divided into two subsets; one set was used to set the rule-based classification and one was used to test the performance of the scheme. RESULTS: In a preliminary clinical study, the implemented detection scheme yielded 98% sensitivity with a false-positive detection rate of less than one false-positive region per image. CONCLUSION: Single-image segmentation methods seem to have high sensitivity in selecting true-positive mass regions in the first stage of a CAD scheme. A multilayer topographic image feature analysis method in the second stage of a CAD scheme has the potential to significantly reduce the false-positive detection rate.     

MR image texture analysis applied to the diagnosis and tracking of Alzheimer''s disease

We assess the value of magnetic resonance (MR) image texture in Alzheimer's disease (AD) both as a diagnostic marker and as a measure of progression. T1-weighted MR scans were acquired from 40 normal controls and 24 AD patients. These were split into a training set (20 controls, 10 AD) and a test set (20 controls, 14 AD). In addition, five control subjects and five AD patients were scanned repeatedly over several years. On each scan a texture feature vector was evaluated over the brain; this consisted of 260 measures derived from the spatial gray-level dependence method. A stepwise discriminant analysis was applied to the training set, to obtain a linear discriminant function. In the test set, this function yielded significantly different values for the control and AD groups (p < 10(-4)) with only small group overlap; a classification rate of 91% was obtained. For the repeatedly scanned control subjects, the median increment in the discriminant function between successive scans of 0.12 was not significantly different from zero (p > 0.05); for the repeatedly scanned AD patients the corresponding median increment of 1.4 was significantly different from zero (p < 0.05). MR image texture may be a useful aid in the diagnosis and tracking of Alzheimer's disease.     

Use of paraffin wax film in MALDI-TOF analysis of DNA

Poor detection limits and strong salt effects are two of the main problems encountered in the matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometric analysis of DNA. This work demonstrates that a probe tip with a paraffin wax film (Parafilm) surface improves the MALDI performance in DNA analysis compared to the commonly used metal surface. First, the use of Parafilm increases the detection sensitivity. It was found that the detection limit achieved with Parafilm was 5 times better than that obtained using stainless steel for a 85mer. More importantly, the Parafilm method could improve detection of larger DNA components in the presence of a large excess of a smaller DNA component or in a DNA mixture. This feature is important to analyses of PCR and sequencing products. Second, we found that the use of Parafilm increased the salt tolerance limits for the 17-, 41-, and 85mers studied in this work and that the salt effect was less sensitive to the DNA size. Third, this method offers other analytical benefits, including producing a more homogeneous coverage of matrix/DNA, adding no extra cost and time to sample preparation, and eliminating the commonly required step for cleaning the probe after analysis. In this paper, we will also present our perspectives on why the use of Parafilm can improve the MALDI-TOF performance in DNA analysis.     

Design of a high-sensitivity classifier based on a genetic algorithm: application to computer-aided diagnosis

A genetic algorithm (GA) based feature selection method was developed for the design of high-sensitivity classifiers, which were tailored to yield high sensitivity with high specificity. The fitness function of the GA was based on the receiver operating characteristic (ROC) partial area index, which is defined as the average specificity above a given sensitivity threshold. The designed GA evolved towards the selection of feature combinations which yielded high specificity in the high-sensitivity region of the ROC curve, regardless of the performance at low sensitivity. This is a desirable quality of a classifier used for breast lesion characterization, since the focus in breast lesion characterization is to diagnose correctly as many benign lesions as possible without missing malignancies. The high-sensitivity classifier, formulated as the Fisher's linear discriminant using GA-selected feature variables, was employed to classify 255 biopsy-proven mammographic masses as malignant or benign. The mammograms were digitized at a pixel size of 0.1 mm x 0.1 mm, and regions of interest (ROIs) containing the biopsied masses were extracted by an experienced radiologist. A recently developed image transformation technique, referred to as the rubber-band straightening transform, was applied to the ROIs. Texture features extracted from the spatial grey-level dependence and run-length statistics matrices of the transformed ROIs were used to distinguish malignant and benign masses. The classification accuracy of the high-sensitivity classifier was compared with that of linear discriminant analysis with stepwise feature selection (LDAsfs). With proper GA training, the ROC partial area of the high-sensitivity classifier above a true-positive fraction of 0.95 was significantly larger than that of LDAsfs, although the latter provided a higher total area (Az) under the ROC curve. By setting an appropriate decision threshold, the high-sensitivity classifier and LDAsfs correctly identified 61% and 34% of the benign masses respectively without missing any malignant masses. Our results show that the choice of the feature selection technique is important in computer-aided diagnosis, and that the GA may be a useful tool for designing classifiers for lesion characterization.     

Time domain analysis of embolic signals can be used in place of high-resolution Wigner analysis when classifying gaseous and particulate emboli

The sample-volume length (SVL) of an embolic signal has previously been used to differentiate between gaseous and particulate emboli and has been calculated using high-resolution Wigner analysis. Although successful, this method of analysis is not widely available to other groups using transcranial Doppler ultrasound (TCD) to classify emboli. The SVL of embolic signals can also be calculated using time domain analysis, which is a far simpler method and potentially available to all TCD users. The aim of this study was to compare the SVL of embolic signals calculated using Wigner analysis and time domain analysis to assess whether or not time domain analysis can replace Wigner analysis to classify emboli. In total, 215 particulate and 100 gaseous emboli were recorded onto digital audiotape and analysed off-line. Two SVLs for each embolic signal were calculated by measuring embolic duration and velocity in the time domain and with Wigner analysis. Receiver operator characteristic (ROC) curves were plotted to assess the optimum SVL threshold for each method, and levels of sensitivity and specificity were defined. The optimum SVL threshold using Wigner analysis was 1.28 cm, yielding 93% sensitivity and 97% specificity. Using time domain analysis, the optimum threshold was 1.12 cm, yielding 90% sensitivity and 96% specificity. The methods were compared statistically (chi2) using their optimum thresholds, and were found not to be statistically different for classifying particles p=0.283) or gaseous emboli (p=0.700). This study has shown that the SVL of embolic signals, used to differentiate particulate from gaseous emboli, can be calculated more simply in the time domain, which yields as accurate results as calculating the SVL using Wigner analysis.     

Computerized image analysis of morphologically transformed and nontransformed Syrian hamster embryo (SHE) cell colonies: application to objective SHE cell transformation assay scoring

We have developed an automated image analysis system that provides comparable classification of morphologically transformed SHE cell colonies to the current visual classification method used in the in vitro SHE cell transformation assay. Visual classification of morphologic transformation in this assay has been shown to accurately predict the carcinogenic potential of chemical, biological and physical agents. The image analysis system is quantitative, based on measuring features of colony color, texture and growth patterns. A linear combination of feature measurements produces a classification process that agrees with visual assessment 93% of the time. All identifiable sources of error are explored and the method is found to be robust in analyzing nearly 500 colonies from a variety of studies performed over a one year period. The high degree of correlation between the visual classification and the objective measurements of the image analysis system validates the reproducibility of the visual scoring process and serves as a basis for automation of the assay.     

A rational approach to detection of significant reflux with duplex Doppler scanning and air plethysmography

PURPOSE: Several techniques are currently available for the detection of venous reflux. We have attempted to determine the relative value and accuracy of available techniques to develop a logical strategy of investigation in reflux venous insufficiency. METHODS: The morphologic distribution of venous incompetence (erect duplex and descending venography); the results of ambulatory venous pressure measurement, venous refilling time, the Valsalva test, and air-plethysmography (venous refilling index, VFI); and the clinical severity were described in 118 consecutive limbs. In an attempt to validate the tests, results were correlated with the clinical severity classification (class 0, n = 34; class 1, n = 42; class 2, n = 11; class 3, n = 31) and with a standardized quantification of reflux (multisegment score) as seen on standing duplex Doppler scanning with rapid deflation cuffs. RESULTS: Twenty-nine percent of limbs with severe venous disease (class 2/3) had pure deep insufficiency, only 6% had pure superficial disease, and the remainder had a combination. A history of previous thrombosis and the presence of posterior tibial vein incompetence were markedly common with ulcer disease (84% and 42%, respectively). The duplex Doppler multisegment score correlated strongly with clinical severity classification (r = 0.97). The venous refilling time and VFI had the highest sensitivity in identifying severe venous disease (class 2/3), and the ambulatory venous pressure had excellent specificity. CONCLUSIONS: For noninvasive determination of reflux, the combination of VFI and duplex scanning not only localized reflux but also separated severe clinical vein disease from mild, with high sensitivity and specificity. Air plethysmography may also provide valuable information regarding calf muscle pump and outflow obstruction.     

Application of artificial neural networks for quantitative analysis of image data in chest radiographs for detection of interstitial lung disease

The authors have developed an automated computeraided diagnostic (CAD) scheme by using artificial neural networks (ANNs) on quantitative analysis of image data. Three separate ANNs were applied for detection of interstitial disease on digitized chest images. The first ANN was trained with horizontal profiles in regions of interest (ROIs) selected from normal and abnormal chest radiographs for distinguishing between normal and abnormal patterns. For training and testing of the second ANN, the vertical output patterns obtained from the 1st ANN were used for each ROI. The output value of the second ANN was used to distinguish between normal and abnormal ROIs with interstitial infiltrates. If the ratio of the number of abnormal ROIs to the total number of all ROIs in a chest image was greater than a specified threshold level, the image was classified as abnormal. In addition, the third ANN was applied to distinguish between normal and abnormal chest images. The combination of the rule-based method and the third ANN also was applied to the classification between normal and abnormal chest images. The performance of the ANNs was evaluated by means of receiver operating characteristic (ROC) analysis. The average Az value (area under the ROC curve) for distinguishing between normal and abnormal cases was 0.976 +/- 0.012 for 100 chest radiographs that were not used in training of ANNs. The results indicate that the ANN trained with image data can learn some statistical properties associated with interstitial infiltrates in chest radiographs.     

Clinically feasible stable isotope technique at a reasonable price: analysis of 13CO2/12CO2-abundance in breath samples with a new isotope selective-nondispersive infrared spectrometer

Up to now, stable isotope analysis of carbon dioxide in breath samples is carried out with sensitive but very expensive and complex isotope ratio mass spectrometry (IRMS). Aiming at a more widespread application of breath tests in gastroenterological diagnostic routine, we tested a newly developed isotope selective non-dispersive infrared spectrometer (NDIRS) in comparison to IRMS. 13C-urea breath tests were performed in 63 patients as the routine screening method for Helicobacter pylori infection. Breath samples at baseline and (15) 30 min after administration of the test solution containing 13C-urea were analysed both by NDIRS and conventional IRMS. The correlation between the delta values of both devices was linear and in good agreement (r = 0.96; p < 0.0001; Y = 1.01 X -0.94). Comparing the delta over baseline-values, the correlation was Y = 1.11 X -0.36 (r = 0.98; p < 0.0001). Referring to the diagnosis of Helicobacter pylori infection with IRMS we calculated a sensitivity of 95.0% and an unchanged specificity (100%) for NDIR analysis. In conclusion, NDIRS appears a promising, easy to operate, and low cost potential alternative to conventional IRMS thus encouraging further detailed investigation and more widespread application of the noninvasive stable isotope technique in breath tests for gastrointestinal function testing.     

Anatomic bases of a vascularized allogenic knee joint transplantation: arterial blood supply of the human knee joint

The reliable identification of tumor cells in populations of atypical cells occurring in body cavity effusions is a well-known diagnostic problem. In order to improve tumor cell detection and to predict disease progression, we developed a cell scoring strategy based on a combination of DNA cytophotometry and immunocytochemistry. For this purpose, morphologically atypical cells obtained from 33 effusion samples were submitted to DNA content analysis and tested for Ber-EP4 immunoreactivity. It turned out that elevated DNA content alone has a low specificity (true negative ratio) and sensitivity (true positive ratio) in predicting disease outcome, whereas Ber-EP4 immunoreactivity alone has a high specificity (100%) but a low sensitivity (56%). In contrast, the use of a scoring system combining the two techniques and relating scores to the previous disease state and the cytomorphology of the atypical cells results in highly specific and sensitive prediction of the disease outcome. We therefore suggest that this approach is a valuable tool for reliably identifying tumor cells in effusions containing populations of cytologically suspect cells.     

The sensitivity and specificity of the ergometric test for the diagnosis of coronary disease

According to the principles of probabilistic analysis, sensitivity and specificity of a diagnostic test are fixed values. Nevertheless, most authors consider them to be inconstant values, specially when applied to the diagnosis of coronary heart disease by exercise stress test. In this paper, we review the basic concepts on sensitivity and specificity of diagnostic tests and try to explain their supposed variability, when related to exercise test, as a function of undue comparison between ST-segment response and the findings of cinecoronariography. Based on the essential difference between coronary heart disease and ischemic heart disease, we demonstrate why such an equivocal comparison can lead to false results of sensitivity and specificity of exercise tests relative to coronary heart disease. As a result, their alleged variability depends most on the prevalence of ischemia throughout the spectrum of coronary heart disease in the studied population. As a matter of fact, unless one can rely on a method as a gold standard for the diagnosis of ischemic heart disease, the real sensitivity and specificity of exercise stress test should be considered as unknown values.     

A multivariate sensory model in glaucoma diagnosis

PURPOSE: To evaluate whether the combination of two psychophysical and two electrophysiological procedures improves diagnostic validity compared with single procedures. METHODS: In a clinical study, 73 patients with glaucoma from the University Eye Hospital in Erlangen and 122 healthy control subjects from the university staff, ranging in age from 19 to 62 years, underwent measurement of temporal contrast sensitivity using a full-field flicker test, spatiotemporal contrast sensitivity, blue-on-yellow visual evoked potential (VEP), and a black-and-white, pattern-reversal electroretinogram. Diagnostic reference criteria included applanation tonometry, optic disc morphometry, and automated perimetry. Sensitivity was determined univariately with a fixed specificity of 80% and in a multivariate approach using logistic regression analysis. The classification rate was estimated using the leaving-one-out method. The correlation with intraocular pressure, visual field defects, and optic nerve defects was determined. RESULTS: Contrast sensitivity measurements and the blue-on-yellow pattern-onset VEP showed comparable sensitivity (85%, 84%, and 85%) with 80% specificity, and a pattern-reversal electroretinogram showed lower sensitivity (64%). The first three methods contributed independent information to a diagnostic score. This score improved sensitivity to 94%, with a specificity of 89%. All procedures moderately correlated with the neuroretinal rim area of the optic disc (r=0.32-0.46). The psychophysical tests showed a higher correlation with visual field defects (r > 0.5) than the electrophysiological tests (r < 0.3). CONCLUSIONS: The multivariate approach substantially increased the diagnostic validity compared with single procedures. This was probably because the diagnostic procedures under investigation tested different aspects of visual function.     

Determination of long-life radiocesiums Cs-134 and Cs-137 in food by gamma-ray spectrometry: summary of collaborative study

PURPOSE: The authors evaluated the ability of a confocal scanning laser ophthalmoscope to detect glaucomatous visual field loss by using their previously described discriminant formula on a prospectively obtained cohort. The relationship of optic disc size to diagnostic classification was also evaluated. METHODS: One eye was chosen randomly from each of 153 subjects. Sixty control eyes had intraocular pressure less than 21 mmHg and normal visual fields; 93 glaucomatous eyes had intraocular pressure greater than 21 mmHg and abnormal visual fields. The optic disc status purposely was not used for classification purposes. All subjects were examined with the Heidelberg Retina Tomograph (HRT; Heidelberg Engineering GMBH, Heidelberg, Germany) and Humphrey Perimeter, program 30-2 (Humphrey Instruments, Inc., San Leandro, CA). Visual fields were considered abnormal by the authors' previously published criteria. The HRT classification used age, adjusted cup shape measure, rim volume, and height variation contour to classify the optic disc as normal or glaucomatous. Then the authors assessed the sensitivity, specificity, and diagnostic precision for the entire group, and for three subsets classified by disc area: disc area less than 2 mm2, between 2 and 3 mm2, and more than 3 mm2. RESULTS: The entire group had a sensitivity, specificity, and diagnostic precision of 74%, 88%, and 80%, respectively. The specificity was 83% when disc area was less than 2 mm2 and improved to 89% when disc area was more than 2 mm2. The sensitivity tended to improve from 65% to 79%, and to 83% if the disc area increased, but the difference was not statistically significant. CONCLUSIONS: In a prospective cohort of patients, the HRT discriminant analysis formula was capable of detecting glaucomatous visual field loss with good precision. Unusually small optic discs continue to present diagnostic difficulties.