Intelligent Imaging: Radiomics and Artificial Neural Networks in Heart Failure

Our previous work with ¹²³iodine meta-iodobenzylguanidine (¹²³I-mIBG) radionuclide imaging among patients with cardiomyopathy reported limitations associated with the prognostic power of global parameters derived from planar imaging [1]. Employing multivariate analysis, we further showed the regional washout associated with territories adjacent to infarcted myocardium obtained from single-photon emission computed tomography imaging (SPECT) yielded superior prognostic power over the other planar and SPECT indices in predicting future cardiac events [1]. The aim of this study was to apply an artificial neural network (Neural Analyser version 2.9.5) to the original data from the same patient cohort to evaluate the most potent prognostic index for future cardiac events among patient with cardiomyopathy.

The original data were reevaluated using an artificial neural network (Neural Analyser version 2.9.5). There were 84 input variables in the original 22 patients from clinical data, electrocardiogram (rest, stress, and continuous ambulatory electrocardiogram recording), transthoracic echocardiography, coronary angiogram, sestamibi myocardial perfusion SPECT, planar and SPECT 123I-mIBG, and genetic and biomarkers, detailed in the previous work. A single binary output was a cardiac event or no cardiac event in the follow-up period.

Following training and validation phases, the optimal number of inputs was determined to be two with a training loss of 0.025 and selection loss <0.001. The final architecture had inputs of a change in left ventricular ejection fraction (Δ > -10%) and 123I-mIBG planar global washout (>30%), two hidden layers of 6 and 1 node, respectively, and a binary output. Using receiver operator characteristics analysis demonstrated an area under the curve of 0.75 correlating to a sensitivity of 100% and specificity of 50%.

The premise that regional washout of 123I-mIBG SPECT from noninfarcted tissue is the best predictor of cardiac events was built on has a sound and logical foundation. By artificial neural network analysis; however, 123I-mIBG planar global washout of >30% was shown to be the best indicator for risk of cardiac event when accompanied by a decline in left ventricular ejection fraction of >10%. Further investigation should be undertaken assessing assimilation into big data and the potential for automated feature extraction from raw image datasets with convolutional neural networks.