Cardiac Pattern Recognition From Spect Images Using Machine Learning Algorithms Publisher



Sabouri M¹ ; Hajianfar G¹ ; Amini M⁴ ; Hosseini Z¹ ; Madadi S¹ ; Ghaedian T³ ; Ghassed M² ; Rastgou F¹ ; Rajabi AB¹ ; Shiri I⁴ ; Zaidi H^{4, 5, 6, 7, 8} Show Affiliations
Source: 2021 IEEE Nuclear Science Symposium and Medical Imaging Conference Record# NSS/MIC 2021 and 28th International Symposium on Room-Temperature Semiconductor Detectors# RTSD 2022 Published:2021

Abstract

Heart failure is a fatal disease that is becoming more prevalent worldwide. Cardiac resynchronization therapy (CRT) treatment is an approach to treat patients with end-stage heart failure. However, since one third of the patients do not respond to this invasive and expensive therapy, response prediction becomes essential for this treatment. Recent studies suggest that patients with a U-shaped left ventricular contraction pattern respond better to CRT treatment. Therefore, our main attempt is to identify these patterns on gated-SPECT myocardial perfusion images (GSPECT MPI) using radiomics and machine learning algorithms to achieve a robust prediction of treatment response. We enrolled 88 patients including 19 patients who underwent CRT, and 69 who did not. In addition to radiomic features, easily accessible clinical features, such as age, sex, QRS complex duration, ejection fraction (EF) and phase analysis data extracted from the quantified gated SPECT (QGS) were analysed. Feature selection was performed with maximum relevant minimum redundancy (MRMR) algorithm. After the feature selection three feature signatures, including a radiomics only, a clinical only and a radiomics + clinical were developed to feed machine learning algorithms. Machine learning techniques included logistic regression (LR), Random Forest (RF), Support Vector Machine (SVM), and XGBoost. The area under the ROC curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE) of all models were reported. The best performance was achieved using the XGB model when applied on the clinical + radiomics feature set (AUC = 0.82). This is followed by that XGB and RF applied to clinical feature signature (AUC = 0.80 and 0.74, respectively). Our results demonstrated the promising potential regarding CRT response prediction with radiomics modelling. © 2021 IEEE.