Identifying coronary stenosis using an image-recognition neural network

Deep learning applications in myocardial perfusion imaging, a systematic review and meta-analysis

Background

Coronary artery disease (CAD) is a leading cause of death worldwide, and the diagnostic process comprises of invasive testing with coronary angiography and non-invasive imaging, in addition to history, clinical examination, and electrocardiography (ECG). A highly accurate assessment of CAD lies in perfusion imaging which is performed by myocardial perfusion scintigraphy (MPS) and magnetic resonance imaging (stress CMR). Recently deep learning has been increasingly applied on perfusion imaging for better understanding of the diagnosis, safety, and outcome of CAD.The aim of this review is to summarise the evidence behind deep learning applications in myocardial perfusion imaging.

Methods

A systematic search was performed on MEDLINE and EMBASE databases, from database inception until September 29, 2020. This included all clinical studies focusing on deep learning applications and myocardial perfusion imaging, and excluded competition conference papers, simulation and animal studies, and studies which used perfusion imaging as a variable with different focus. This was followed by review of abstracts and full texts. A meta-analysis was performed on a subgroup of studies which looked at perfusion images classification. A summary receiver-operating curve (SROC) was used to compare the performance of different models, and area under the curve (AUC) was reported. Effect size, risk of bias and heterogeneity were tested.

Results

46 studies in total were identified, the majority were MPS studies (76%). The most common neural network was convolutional neural network (CNN) (41%). 13 studies (28%) looked at perfusion imaging classification using MPS, the pooled diagnostic accuracy showed AUC = 0.859. The summary receiver operating curve (SROC) comparison showed superior performance of CNN (AUC = 0.894) compared to MLP (AUC = 0.848). The funnel plot was asymmetrical, and the effect size was significantly different with p value < 0.001, indicating small studies effect and possible publication bias. There was no significant heterogeneity amongst studies according to Q test (p = 0.2184).

Conclusion

Deep learning has shown promise to improve myocardial perfusion imaging diagnostic accuracy, prediction of patients' events and safety. More research is required in clinical applications, to achieve better care for patients with known or suspected CAD.

Toward automatic detection of vessel stenoses in cerebral 3D DSA volumes

Vessel diseases are a very common reason for permanent organ damage, disability and death. This fact necessitates further research for extracting meaningful and reliable medical information from the 3D DSA volumes. Murray's law states that at each branch point of a lumen-based system, the sum of the minor branch diameters each raised to the power x, is equal to the main branch diameter raised to the power x. The principle of minimum work and other factors like the vessel type, impose typical values for the junction exponent x. Therefore, deviations from these typical values may signal pathological cases. In this paper, we state the necessary and the sufficient conditions for the existence and the uniqueness of the solution for x. The second contribution is a scale- and orientation- independent set of features for stenosis classification. A support vector machine classifier was trained in the space of these features. Only one branch was misclassified in a cross validation on 23 branches. The two contributions fit into a pipeline for the automatic detection of the cerebral vessel stenoses.

Artificial intelligence for diagnostic purposes: principles, procedures and limitations

BACKGROUND

Back propagation (BP) artificial neural networks are a distribution-free method for data analysis based on layers of artificial neurons that transduce imputed information. It has been recognized as having a number of advantages compared to traditional methods including the possibility to process imperfect data, and complex non-linear data. The objective of this study was to review the principles, procedures, and limitations of BP artificial neural networks for a non-mathematical readership.

METHODS

A real data sample of weight, height and measured body surface area from 90 individuals was used as an example. SPSS 17.0 with neural network add-on was used for the analysis. The predicted body surface from a two hidden layer BP neural network was compared to the body surface calculated by the Haycock equation.

RESULTS

Both the predicted values from the neural network and from the Haycock equation were close to the measured values. A linear regression analysis with neural network as predictor produced an r(2)-value of 0.983, while the Haycock equation produced an r(2)-value of 0.995 (r(2)>0.95 is a criterion for accurate diagnostic testing).

CONCLUSIONS

BP neural networks may, sometimes, predict clinical diagnoses with accuracies similar to those of other methods. However, traditional statistical procedures, such as regression analyses need to be added for testing their accuracies against alternative methods. Nonetheless, BP neural networks have great potential through their ability to learn by example instead of learning by theory.

Neural network predictions of significant coronary artery stenosis in men

OBJECTIVE

A neural network system was designed to predict whether coronary arteriography on a given patient would reveal any occurrence of significant coronary stenosis (>50%), a degree of stenosis which often leads to coronary intervention.

METHODOLOGY

A dataset of 2004 records from male cardiology patients was derived from a national cardiac catheterization database. The catheterizations selected for analysis from the database were first-time and elective, and they were precipitated by chest pain. Eleven patient variables were used as inputs in an artificial neural network system. The network was trained on the earliest 902 records in the dataset. The next 902 records formed a cross-validation file, which was used to optimize the training. A third file composed of the next 100 records facilitated the choice of a cutoff number between 0 and 1. The cutoff number was applied to the last 100 records, which comprised a test file.

RESULTS

When a cutoff of 0.25 was compared to the network outputs of all 100 records in the test file, 12 of 46 (specificity=26%) patients without significant stenosis had outputs<or=0.25, but all patients with significant stenosis had outputs>0.25 (sensitivity=100%). Therefore, the network identified a fraction of the patients in the test file who did not have significant coronary artery stenosis, while at the same time the network identified all of the patients in the test file who had significant stenosis capable of causing chest pain.

CONCLUSION

Artificial neural networks may be helpful in reducing unnecessary cardiac catheterizations.

A review of evidence of health benefit from artificial neural networks in medical intervention

The purpose of this review is to assess the evidence of healthcare benefits involving the application of artificial neural networks to the clinical functions of diagnosis, prognosis and survival analysis, in the medical domains of oncology, critical care and cardiovascular medicine. The primary source of publications is PUBMED listings under Randomised Controlled Trials and Clinical Trials. The rĵle of neural networks is introduced within the context of advances in medical decision support arising from parallel developments in statistics and artificial intelligence. This is followed by a survey of published Randomised Controlled Trials and Clinical Trials, leading to recommendations for good practice in the design and evaluation of neural networks for use in medical intervention.

Improving cephalogram analysis through feature subimage extraction

An MLP with a GA was proposed to extract feature subimages containing orthodontic landmarks. Simulated images and cephalograms were used to investigate its performance in comparison with the cross-correlation method. From the results of simulated image containing shapes with different geometrical conditions, it was shown that the fault tolerance of the MLP for rotation, scaling, brightness variety, and other anomalous deformations is good enough to overcome the clinical application problems. It was also shown that the stability, accuracy, and speed of this proposed algorithm are very promising. Moreover, the performance of the MLP can be significantly improved by collecting, more "representative" false patterns. The GA is a good approach to speed up the process of feature subimage extraction based on the fitness evaluated using the MLP.