Skip to main content
Fig. 2 | Journal of Cardiovascular Magnetic Resonance

Fig. 2

From: Neural-network classification of cardiac disease from 31P cardiovascular magnetic resonance spectroscopy measures of creatine kinase energy metabolism

Fig. 2

Schematic of the neural network used for classifying heart disease, heart failure (HF) type and New York Heart Association (NYHA) HF class, based on noninvasive measures of creatine kinase (CK) metabolism ([PCr], [ATP], kf and CK flux), as implemented (HCM ± = hypertrophy with/without HF; MI = anterior myocardial infarction). Each grey dot represents a node with an input, a weight and a bias term. For the first layer, there are 25 parameters (4 weights plus 1 bias, times 5 nodes); the second layer has 24 (5 inputs times 4 nodes, plus 4 biases, 1 per node); the third ‘Softmax’ layer has 20 parameters (4 times 4, plus 4). The total number of parameters is thus 69 or about 10% of the 664 samples (166 times 4) used in the initial training for disease type (Table 2. This falls to 4% of the 4 × 388 = 1552 points when the synthetic data in Table 6 are included). The maximum number of training epochs was 8000

Back to article page