Use of neural networks to improve quality control of interpretations in myocardial perfusion imaging.

BACKGROUND: The aim of this study was to explore the feasibility of using a technique based on artificial neural networks for quality assurance of image reporting. The networks were used to identify potentially suboptimal or erroneous interpretations of myocardial perfusion scintigrams (MPS). METHODS: Reversible perfusion defects (ischaemia) in each of five myocardial regions, as interpreted by one experienced nuclear medicine physician during his daily routine of clinical reporting, were assessed by artificial neural networks in 316 consecutive patients undergoing stress/rest 99mTc-sestamibi myocardial perfusion scintigraphy. After a training process, the networks were used to select the 20 cases in each region that were more likely to have a false clinical interpretation. These cases, together with 20 control cases in which the networks detected no likelihood of false clinical interpretation, were presented in random order to a group of three experienced physicians for a consensus re-interpretation; no information regarding clinical or neural network interpretations was provided to the re-evaluation panel. RESULTS: The clinical interpretation and the re-evaluation differed in 53 of the 200 cases. Forty-six of the 53 cases (87%) came from the group selected by the neural networks, and only seven (13%) were control cases (P < 0.001). The disagreements between clinical routine interpretation by an experienced nuclear medicine expert and artificial networks were related to small and mild perfusion defects and localization of defects. CONCLUSION: The results demonstrate that artificial neural networks can identify those myocardial perfusion scintigrams that may have suboptimal image interpretations. This is a potentially highly cost-effective technique, which could be of great value, both in daily practice as a clinical decision support tool and as a tool in quality assurance.

A decision support system improves the interpretation of myocardial perfusion imaging.

PURPOSE: The aim of this study was to investigate the influence of a computer-based decision support system (DSS) on performance and inter-observer variability of interpretations regarding ischaemia and infarction in myocardial perfusion scintigraphy (MPS). METHODS: Seven physicians independently interpreted 97 MPS studies, first without and then with the advice of a DSS. Four physicians had long experience and three had limited experience in the interpretation of MPS. Each study was interpreted regarding myocardial ischaemia and infarction in five myocardial regions. The patients had undergone a gated MPS using a 2-day stress/gated rest (99m)Tc sestamibi protocol. The gold standard used was the interpretations made by one experienced nuclear medicine specialist on the basis of all available clinical and image information. RESULTS: The sensitivity for ischaemia of the seven readers increased from 81% without the DSS to 86% with the DSS (p = 0.01). The increase in sensitivity was higher for the three inexperienced physicians (9%) than for the four experienced physicians (2%). There was no significant change in specificity between the interpretations. The interpretations of ischaemia made with the advice of the DSS showed less inter-observer variability than those made without advice. CONCLUSION: This study shows that a DSS can improve performance and reduces the inter-observer variability of interpretations in myocardial perfusion imaging. Both experienced and, especially, inexperienced physicians can improve their interpretation with the advice from such a system.

EANM/ESC guidelines for radionuclide imaging of cardiac function.

Radionuclide imaging of cardiac function represents a number of well-validated techniques for accurate determination of right (RV) and left ventricular (LV) ejection fraction (EF) and LV volumes. These first European guidelines give recommendations for how and when to use first-pass and equilibrium radionuclide ventriculography, gated myocardial perfusion scintigraphy, gated PET, and studies with non-imaging devices for the evaluation of cardiac function. The items covered are presented in 11 sections: clinical indications, radiopharmaceuticals and dosimetry, study acquisition, RV EF, LV EF, LV volumes, LV regional function, LV diastolic function, reports and image display and reference values from the literature of RVEF, LVEF and LV volumes. If specific recommendations given cannot be based on evidence from original, scientific studies, referral is given to "prevailing or general consensus". The guidelines are designed to assist in the practice of referral to, performance, interpretation and reporting of nuclear cardiology studies for the evaluation of cardiac performance.

EANM/ESC procedural guidelines for myocardial perfusion imaging in nuclear cardiology.

The European procedural guidelines for radionuclide imaging of myocardial perfusion and viability are presented in 13 sections covering patient information, radiopharmaceuticals, injected activities and dosimetry, stress tests, imaging protocols and acquisition, quality control and reconstruction methods, gated studies and attenuation-scatter compensation, data analysis, reports and image display, and positron emission tomography. If the specific recommendations given could not be based on evidence from original, scientific studies, we tried to express this state-of-art. The guidelines are designed to assist in the practice of performing, interpreting and reporting myocardial perfusion SPET. The guidelines do not discuss clinical indications, benefits or drawbacks of radionuclide myocardial imaging compared to non-nuclear techniques, nor do they cover cost benefit or cost effectiveness.

Operator dependent variability in quantitative analysis of myocardial perfusion images.

The purpose of this study was to evaluate variability in the quantification of myocardial perfusion images obtained by a group of experienced operators using two widely used programs. The Cedars Emory quantitative analysis program (CEqual) was used to quantify the size of perfusion defects and the Cedars-Sinai quantitative gated single-photon emission tomography program was used to quantify left ventricular function. Five patients with reversible apical defects, five with fixed apical defects and three patients with normal perfusion were selected. Eight experienced medical laboratory technologists processed the studies from raw projection data. The manual steps consisted of defining two alignment axes parallel to the long axis of the left ventricle, and for the CEqual program selecting apex and base in the short axis slices in the rest and stress studies. Wide variability between the operators in the quantification of reversibility could be seen in all three vascular territories. A range >10% was found in at least one vascular territory for nine of the 13 patients. The differences in left ventricular ejection fraction (LVEF) between operators were <5% for all 13 patients. The large variability in the quantification of reversible apical perfusion defects may influence the clinical interpretation and cause false conclusions. In contrast, inter-operator variability for the quantification of the LVEF was low.

The optimal reference population for cardiac normality in myocardial SPET in the detection of coronary artery stenoses: patients with normal coronary angiography or subjects with low likelihood of coronary artery disease?

Both subjects with a low likelihood for coronary artery disease (CAD) and patients with normal findings on coronary angiography have been used as reference populations in non-invasive stress testing, including myocardial perfusion scintigraphy. Both of these criteria of normality--low likelihood of CAD and normal coronary angiography--have been criticised, and consensus on this issue is lacking. The aim of this study was to compare two different reference populations by testing the performance of artificial neural networks designed to interpret myocardial scintigrams. The networks were trained on myocardial perfusion scintigrams from 87 patients with angiographically documented CAD and on studies from one of two different reference groups: 48 patients with no signs of CAD based on angiography or 128 healthy volunteers with a likelihood for CAD <5%. The performance of the two different networks was then tested using scintigrams from a separate test group of 68 patients. Coronary angiography was used as the gold standard in this group. The network trained on patients with no signs of CAD based on angiography showed an area under the receiver operating characteristic (ROC) curve of 93%. The ROC area for the network trained on healthy volunteers was 72%, and this difference was statistically significant (P=0.03). The results of this study using artificial neural networks suggest that normal angiography should be preferred as the reference standard in myocardial scintigraphy when a patient is examined for CAD prior to possible angiography. Whether the same is true for other indications, e.g. in prognostic evaluation, is unknown.

Usefulness of serial electrocardiograms for diagnosis of acute myocardial infarction.

The purpose of this study was to determine whether the automated detection of acute myocardial infarction (AMI) by utilizing artificial neural networks was improved by using a previous electrocardiogram (ECG) in addition to the current ECG. A total of 4,691 ECGs were recorded from patients admitted to an emergency department due to suspected AMI. Of these, 902 ECGs, in which diagnoses of AMI were later confirmed, formed the study group, whereas the remaining 3,789 ECGS comprised the control group. For each ECG recorded, a previous ECG of the same patient was selected from the clinical electrocardiographic database. Artificial neural networks were then programmed to detect AMI based on either the current ECG only or on the combination of the previous and the current ECGs. On this basis, 3 assessors--a neural network, an experienced cardiologist, and an intern--separately classified the ECGs of the test group, with and without access to the previous ECG. The detection performance, as measured by the area under the receiver operating characteristic curve, showed an increase for all assessors with access to previous ECGs. The neural network improved from 0.85 to 0.88 (p = 0.02), the cardiologist from 0.79 to 0.81 (p = 0.36), and the intern from 0.71 to 0.78 (p <0.001). Thus, the performance of a neural network, detecting AMI in an ECG, is improved when a previous ECG is used as an additional input.

An independent evaluation of a new method for automated interpretation of lung scintigrams using artificial neural networks.

The purpose of this study was to evaluate a new automated method for the interpretation of lung perfusion scintigrams using patients from a hospital other than that where the method was developed, and then to compare the performance of the technique against that of experienced physicians. A total of 1,087 scintigrams from patients with suspected pulmonary embolism comprised the training group. The test group consisted of scintigrams from 140 patients collected in a hospital different to that from which the training group had been drawn. An artificial neural network was trained using 18 automatically obtained features from each set of perfusion scintigrams. The image processing techniques included alignment to templates, construction of quotient images based on the perfusion/template images, and finally calculation of features describing segmental perfusion defects in the quotient images. The templates represented lungs of normal size and shape without any pathological changes. The performance of the neural network was compared with that of three experienced physicians who read the same test scintigrams according to the modified PIOPED criteria using, in addition to perfusion images, ventilation images when available and chest radiographs for all patients. Performances were measured as area under the receiver operating characteristic curve. The performance of the neural network evaluated in the test group was 0.88 (95% confidence limits 0.81-0.94). The performance of the three experienced experts was in the range 0.87-0.93 when using the perfusion images, chest radiographs and ventilation images when available. Perfusion scintigrams can be interpreted regarding the diagnosis of pulmonary embolism by the use of an automated method also in a hospital other than that where it was developed. The performance of this method is similar to that of experienced physicians even though the physicians, in addition to perfusion images, also had access to ventilation images for most patients and chest radiographs for all patients. These results show the high potential for the method as a clinical decision support system.

Changes in high-frequency QRS components are more sensitive than ST-segment deviation for detecting acute coronary artery occlusion.

OBJECTIVES: This study describes changes in high-frequency QRS components (HF-QRS) during percutaneous transluminal coronary angioplasty (PTCA) and compares the ability of these changes in HF-QRS and ST-segment deviation in the standard 12-lead electrocardiogram (ECG) to detect acute coronary artery occlusion. BACKGROUND: Previous studies have shown decreased HF-QRS in the frequency range of 150-250 Hz during acute myocardial ischemia. It would be important to know whether the high-frequency analysis could add information to that available from the ST segments in the standard ECG. METHODS: The study population consisted of 52 patients undergoing prolonged balloon occlusion during PTCA. Signal-averaged electrocardiograms (SAECG) were recorded prior to and during the balloon inflation. The HF-QRS were determined within a bandwidth of 150-250 Hz in the preinflation and inflation SAECGs. The ST-segment deviation during inflation was determined in the standard frequency range. RESULTS: The sensitivity for detecting acute coronary artery occlusion was 88% using the high-frequency method. In 71% of the patients there was ST elevation during inflation. If both ST elevation and depression were considered, the sensitivity was 79%. The sensitivity was significantly higher using the high-frequency method, p<0.002, compared with the assessment of ST elevation. CONCLUSIONS: Acute coronary artery occlusion is detected with higher sensitivity using high-frequency QRS analysis compared with conventional assessment of ST segments. This result suggests that analysis of HF-QRS could provide an adjunctive tool with high sensitivity for detecting acute myocardial ischemia.

Clustering ECG complexes using hermite functions and self-organizing maps.

An integrated method for clustering of QRS complexes is presented which includes basis function representation and self-organizing neural networks (NN's). Each QRS complex is decomposed into Hermite basis functions and the resulting coefficients and width parameter are used to represent the complex. By means of this representation, unsupervised self-organizing NN's are employed to cluster the data into 25 groups. Using the MIT-BIH arrhythmia database, the resulting clusters are found to exhibit a very low degree of misclassification (1.5%). The integrated method outperforms, on the MIT-BIH database, both a published supervised learning method as well as a conventional template cross-correlation clustering method.

Scandinavian test of artificial neural network for classification of myocardial perfusion images.

Artificial neural networks are systems of elementary computing units capable of learning from examples. They have been applied to automated interpretation of myocardial perfusion images and have been shown to perform even better than experienced physicians. It has been shown that physicians interpreting myocardial perfusion images benefit from the advice of such networks. These networks have been developed and validated in the same hospital. However, widespread use of neural networks will only take place if the networks can maintain a high accuracy in other hospitals, i.e. hospitals using different gamma cameras, different acquisition techniques, different study protocols, etc. The purpose of this study was to develop a neural network in one hospital and test it in another. An artificial neural network was trained to detect coronary artery disease using myocardial perfusion scintigrams from 135 patients at a Swedish hospital. Thereafter, this network was tested using scintigrams from 68 patients at a Danish hospital and compared to six criteria based on expert physician analysis and quantitative analysis by the CEqual program. The sensitivity of the network was significantly higher than that of one of the physician criteria (0. 92 versus 0.71) and two of the CEqual-based criteria (0.94 versus 0. 63 and 0.96 versus 0.65) compared at equal specificities. It was concluded that an artificial neural network can maintain high accuracy in a hospital other than the one where it was developed.

Automated interpretation of ventilation-perfusion lung scintigrams for the diagnosis of pulmonary embolism using artificial neural networks.

The purpose of this study was to develop a completely automated method for the interpretation of ventilation-perfusion (V-P) lung scintigrams used in the diagnosis of pulmonary embolism. An artificial neural network was trained for the diagnosis of pulmonary embolism using 18 automatically obtained features from each set of V-P scintigrams. The techniques used to process the images included their alignment to templates, the construction of quotient images based on the ventilation and perfusion images, and the calculation of measures describing V-P mismatches in the quotient images. The templates represented lungs of normal size and shape without any pathological changes. Images that could not be properly aligned to the templates were detected and excluded automatically. After exclusion of those V-P scintigrams not properly aligned to the templates, 478 V-P scintigrams remained in a training group of consecutive patients with suspected pulmonary embolism, and a further 87 V-P scintigrams formed a separate test group comprising patients who had undergone pulmonary angiography. The performance of the neural network, measured as the area under the receiver operating characteristic curve, was 0.87 (95% confidence limits 0.82-0.92) in the training group and 0.79 (0.69-0.88) in the test group. It is concluded that a completely automated method can be used for the interpretation of V-P scintigrams. The performance of this method is similar to others previously presented, whereby features were extracted manually.

Spatial, individual, and temporal variation of the high-frequency QRS amplitudes in the 12 standard electrocardiographic leads.

BACKGROUND: Analysis of high-frequency QRS amplitudes (HF-QRS) may provide an additional diagnostic tool in patients with heart disease, but the basic properties of these waveforms have not been sufficiently investigated. This study describes the spatial, individual, and temporal variation at rest of HF-QRS recorded with the 12 standard electrocardiographic leads in patients with ischemic heart disease. METHODS AND RESULTS: Two consecutive electrocardiographic recordings from 67 patients were signal averaged and analyzed within a bandwidth of 150 to 250 Hz. The HF-QRS values were expressed as root mean square values. There was a spatial variation in HF-QRS among the 12 leads, with higher amplitudes in V(2) through V(4), II, aVF, and III. The individual variation among the patients was large for all leads. The sum of the HF-QRS for all 12 leads in each patient ranged from 20 to 75 microV (mean 36 +/- 11 microV). The mean of the temporal variation in HF-QRS for all 12 leads between the 2 recordings was only 0.10 +/- 0. 09 microV. CONCLUSIONS: Because of the large individual variation, analysis of HF-QRS is probably most applicable in monitoring situations when it is possible to track changes in a patient over time. The temporal variation in HF-QRS at rest is small, both in patients with and those without prior myocardial infarction.

Myocardial SPET: artificial neural networks describe extent and severity of perfusion defects.

Artificial neural networks are computer programs that learn from examples. They have been successfully used to detect coronary artery disease from myocardial perfusion images. The purpose of the present study was to develop neural networks that could classify myocardial scintigrams regarding reversibility, localization, severity and extent of perfusion defects. Rest/exercise technetium-99m sestamibi scintigrams from 338 patients were studied. The classifications of two experts were employed as the gold standard. Artificial neural networks were trained to classify both reversible (ischaemia) and non-reversible (infarct) defects in three vascular territories, corresponding to the main coronary arteries. The extent (small or large) and severity (mild or severe) of the defects were described by the networks. After the training process, separate test sets were used to compare the neural networks with one of the experts who reclassified the scintigrams two months later. The neural networks made correct classifications in 71% of the test cases and the human expert in 70% (P=0.10). It was concluded that artificial neural networks can be trained to make clinical interpretations of myocardial perfusion scintigrams. The results indicate that networks can assist physicians in achieving correct interpretations and thereby improve the diagnostic accuracy of medical imaging.

A confident decision support system for interpreting electrocardiograms.

Computer-aided interpretation of electrocardiograms (ECGs) is widespread but many physicians hesitate to rely on the computer, because the advice is presented without information about the confidence of the advice. The purpose of this work was to develop a method to validate the advice of a computer by estimating the error of an artificial neural network output. A total of 1249 ECGs, recorded with computerized electrocardiographs, on patients who had undergone diagnostic cardiac catheterization were studied. The material consisted of two groups, 414 patients with and 835 without anterior myocardial infarction. The material was randomly divided into three data sets. The first set was used to train an artificial neural network for the diagnosis of anterior infarction. The second data set was used to calculate the error of the network outputs. The last data set was used to test the network performance and to estimate the error of the network outputs. The performance of the neural network, measured as the area under the receiver operating characteristic (ROC) curve, was 0.887 (0.845-0.922). The 25% test ECGs with the lowest error estimates had an area under the ROC curve as high as 0.995 (0.982-1.000), i.e. almost all of these ECGs were correctly classified. Neural networks can therefore be trained to diagnose myocardial infarction and to signal when the advice is given with great confidence or when it should be considered more carefully. This method increases the possibility that artificial neural networks will be accepted as reliable decision support systems in clinical practice.

Improved classifications of myocardial bull's-eye scintigrams with computer-based decision support system.

UNLABELLED: In a recent study, artificial neural networks were trained to detect coronary artery disease using scintigraphic data as input. The performance of the networks was better than that of human experts using coronary angiography as a gold standard. In clinical practice, this type of neural networks will not take over the decision-making process from the physician but will assist by proposing an interpretation of the scintigram. The purpose of this study was to assess the influence of such decision support on the interpretations of the physicians. METHODS: A population of 135 patients who had undergone both myocardial 99mTc-sestamibi rest/stress scintigraphy and coronary angiography within a 3-mo period was studied. An image set consisting of the bull's-eye rest, stress, difference and quote images was constructed for each patient. Three experienced physicians independently classified all image sets regarding the presence and/or absence of coronary artery disease in two vascular territories using a four-grade scale. The physicians classified the image sets twice with and twice without the advice of artificial neural networks. RESULTS: The joint evaluation of the three physicians showed significantly improved performance with decision support, measured as increases in the areas under the receiver operating characteristic curves from 0.65 to 0.70 (P = 0.018) and from 0.79 to 0.82 (P = 0.006) for two vascular territories. Furthermore, the joint evaluation showed significantly less intraobserver and interobserver variability with decision support. CONCLUSION: Physicians classifying myocardial bull's-eye images benefit from the advice of artificial neural networks. These results show the high potential for neural networks as clinical decision support systems.

Recognition of lead reversals in pediatric electrocardiograms.

We have trained artificial neural networks to detect right/left arm lead reversals in pediatric electrocardiograms with a performance significantly higher than that of currently used methods. We believe that this type of neural network can be a valuable type of quality control in pediatric electrocardiography.

Scintigraphic diagnosis of coronary artery disease: myocardial bull's-eye images contain the important information.

The bull's-eye image, also called polar map image, has been developed as an important display for the visual and quantitative analysis of myocardial perfusion scintigrams. Quantitative analysis can be performed for example by comparing areas in the bull's-eye image with normal limits or by processing it using artificial neural networks. The usefulness of such methods is highly dependent on the information content of the bull's-eye image. The purpose of this study was to investigate whether there is more diagnostically important information in a set consisting of the myocardial bull's-eye image plus tomographic slice image than in the bull's-eye image alone. A population of 135 patients who had undergone both myocardial scintigraphy and coronary angiography, with no more than 3 months elapsing between the two examinations, was studied retrospectively. Four experienced observers independently classified visually all scintigrams regarding the presence/absence of coronary artery disease in two vascular territories using a four-grade scale. The observers classified the scintigrams once viewing bull's-eye images only, and once viewing tomographic slices and bull's-eye images. Coronary angiography was used as gold standard. The classifications were evaluated using the areas under the receiver operating characteristics (ROC) curves. The classifications based on bull's-eye images only were slightly more accurate than those based on tomographic slices and bull's-eye images in one of the two vascular territories (ROC areas of 0.66 vs. 0.64). The opposite relationship was found in the other vascular territory (0.78 vs. 0.81). None of the differences was statistically significant. In conclusion, the diagnostically important information for the diagnosis of coronary artery disease by myocardial perfusion scintigraphy is present in the bull's-eye image.

Myocardial infarction--pinpointing the key indicators in the 12-lead ECG using data mining.

In this paper we describe how data mining techniques were used in order to pinpoint the key indicators for myocardial infarction in the electrocardiogram (ECG) by determining existing trends in a large data set. In order to provide a test bed for the data mining techniques a data mining tool was developed so that the effectiveness of various data mining techniques could be determined. The material consisted of 2730 ECGs recorded at an emergency department. A total of 517 ECGs were recorded on patients suffering acute myocardial infarction. The remaining ECGs were defined as control ECGs. A subset of the material was used to train the data mining tool. After training, the data mining tool was able to pinpoint the key ECG indicators for myocardial infarction in the test set (duration and amplitude of the Q wave and R duration in lead V2) and successfully determine which patients had suffered a heart attack.

Intelligent computer reporting 'lack of experience': a confidence measure for decision support systems.

The purpose of this study was to explore the feasibility of developing artificial neural networks that are able to provide confidence measures for their diagnostic advice. Computer-aided decision making can improve physician performance, but many physicians hesitate to use these 'black boxes'. If we are to rely upon decision support systems for such tasks as medical diagnosis it is essential that the computers indicate when the advice given is based on experience, i.e. give a confidence measure. An artificial neural network was trained to diagnose healed anterior myocardial infarction and to indicate 'lack of experience' when test electrocardiograms were different from the electrocardiograms of the training set. A database of 1249 electrocardiograms from patients who had undergone cardiac catheterization was used to train and test the neural network. Thereafter, the ability of the network to indicate 'lack of experience' was assessed using 100 left bundle branch block electrocardiograms, an electrocardiographic pattern that was excluded from the training set. The network indicated that 83% of the left bundle branch block electrocardiograms and 1% of the test electrocardiograms from catheterized patients were different from the electrocardiograms of the training set. All but one of the left bundle branch block electrocardiograms would otherwise be falsely classified as anterior myocardial infarction by the network. Artificial neural networks can be trained to indicate 'lack of experience', and this ability increases the possibility for neural networks to be accepted as reliable decision support systems in clinical practice.