Maximizing the Predictive Power of Radiomic Features in SPECT Images: A Comparative Study of Reconstruction Algorithms Using Machine Learning

Cardiovascular diseases are the leading causes of death worldwide. SPECT imaging is an effective, non-invasive method that can be used to assess and study cardiac diseases. Using quantitative features of radiomics, more useful information can be extracted from the SPECT images. Considering that different reconstruction settings affect radiomic features, we aim to select optimal reconstruction settings for SPECT imaging to extract the most effective features. To investigate this, 15 different reconstruction settings were performed on stress and rest SPECT images of 200 patients. Subsequently, the left ventricles of all patients were segmented, and radiomic features were extracted from them. Following this, modeling was performed on the 15 datasets, using 3 feature selections and 30 machine learning methods, and the features extracted from two sets of images reconstructed by 11th reconstruction (#11) (Iterative, Butterworth, Cutoff=0.40, Order=5, (Iteration, subset)=(4-4)) and 10th reconstruction methods (#10) (Iterative, Butterworth, Cutoff=0.35, Order=5, (Iteration, subset)=(4-4)) were shown to outperform results from other reconstructions. Thus, all three top models that selected features using the MRMR and ANOVA methods came from the #11. A model related to the #11 and two related to the #10 were among the top three models when employing the RFE feature selection approach. Overall, the best model obtained in this study was SVM-RFE (AUC:0.97, accuracy: 0.96, balanced accuracy: 0.97, F1 Score: 0.96).