Classification of ischemia from myocardial polar maps in 15O-H2O cardiac perfusion imaging using a convolutional neural network

Abstract

We implemented a two-dimensional convolutional neural network (CNN) for classification of polar maps extracted from Carimas (Turku PET Centre, Finland) software used for myocardial perfusion analysis. 138 polar maps from ¹⁵O-H₂O stress perfusion study in JPEG format from patients classified as ischemic or non-ischemic based on finding obstructive coronary artery disease (CAD) on invasive coronary artery angiography were used. The CNN was evaluated against the clinical interpretation. The classification accuracy was evaluated with: accuracy (ACC), area under the receiver operating characteristic curve (AUC), F1 score (F1S), sensitivity (SEN), specificity (SPE) and precision (PRE). The CNN had a median ACC of 0.8261, AUC of 0.8058, F1S of 0.7647, SEN of 0.6500, SPE of 0.9615 and PRE of 0.9286. In comparison, clinical interpretation had ACC of 0.8696, AUC of 0.8558, F1S of 0.8333, SEN of 0.7500, SPE of 0.9615 and PRE of 0.9375. The CNN classified only 2 cases differently than the clinical interpretation. The clinical interpretation and CNN had similar accuracy in classifying false positives and true negatives. Classification of ischemia is feasible in ¹⁵O-H₂O stress perfusion imaging using JPEG polar maps alone with a custom CNN and may be useful for the detection of obstructive CAD.

Figure 1

An example of an (a) original polar map JPEG image before cropping and resizing to smaller dimensions. After cropping and resizing, (b) only the polar map and the superimposed regional markings signifying LAD, LCX and RCA remain in a rectangular shape of 256 × 256 pixels. All polar maps were scaled between 0 and 3.5 ml/g/min before image export, where red and yellow color indicate normal stress MBF, whereas green and blue color indicate reduced stress MBF.

Figure 2

The 2D CNN architecture used in this study. The input polar map is a 256 × 256 RGB image. Each 2D convolutional layer contains a 3 × 3 kernel with stride 2 × 2 and ReLU as activation function. Filter sizes are increased gradually from 12, 16, 32 to 64. Each convolutional layer is followed by a max pooling layer with a window size of 2 × 2. Thus, the output of the last convolution and max-pooling layer is reduced to 1 × 1 × 64, which is given as an input to the flatten layer. The last layers contain a flatten layer of size 64 followed by two dense layer of sizes 512 and 128, using ReLU as activation function. The output layer is a dense layer with 1 output (0 or 1) with a sigmoid function. The input size of each convolutional layer and the flatten layer is given above each layer. For the fully connected layers, the output size is given. conv = 2D convolutional layer, flatten = flatten layer, fc = fully connected layer.

Figure 3

Workflow diagram of the analysis pipeline used to select the optimum parameters for the CNN and to evaluate the performance of the final model. The training, validation and test steps were repeated 100 times for the CNN to evaluate the model stability.

Figure 4

CNN classification accuracy with (a) training (b) validation and (c) test data. ACC = accuracy, AUC = area under the receiver operating characteristic curve, F1S = F1 score, SEN = sensitivity, SPE = specificity, PRE = precision.

Figure 5

Visualization of polar maps classified as true positive (TP), true negative (TN), false positive (FP) and false negative (FN) from the best classification result with the CNN (accuracy 0.8696, matching the clinical interpretation).

Figure 6

CNN classification performance from 100 runs of the network versus the clinical interpretation. The best classification results of the CNN match the performance of clinical interpretation ACC = accuracy, AUC = area under the receiver operating characteristic curve, F1S = F1 score, SEN = sensitivity, SPE = specificity, PRE = precision.

Figure 7

Decision curve with net benefit for all threshold probabilities from 0 to 100% for CNN median over 100 runs and the clinical interpretation using the test data. Only positive values in the range of 0 to 1 for the net benefit are shown.