Detection of myocardial infarction by postmortem CT combined with radiomics-based machine-learning analysis

To determine whether quantitative radiomic features extracted from noncontrast postmortem computed tomography (PMCT) can reliably detect autopsy-confirmed myocardial infarction (MI), and to evaluate the diagnostic performance of machine-learning models for automated MI detection.

We retrospectively analyzed 106 PMCT examinations (55 MI cases, 51 controls). The left ventricular myocardium was semiautomatically segmented, and radiomic features were extracted. After reproducibility filtering (intraclass correlation coefficient ≥ 0.75), redundancy reduction based on Spearman correlation (ρ > 0.8), and feature selection were performed within a nested stratified cross-validation framework, including no additional selection (N.S.), random forest (RF)-based ranking, and recursive feature elimination with logistic regression (RFE-LR). Seven supervised machine-learning classifiers were trained and evaluated using area under the receiver operating characteristic curve (AUROC) and other metrics.

Logistic regression with N.S. achieved the highest performance, with a mean AUROC of 0.817 ± 0.12 (mean ± standard deviation), with accuracy 0.746 ± 0.05, sensitivity 0.764 ± 0.14, and specificity 0.724 ± 0.13. Support vector machines and RF with N.S. showed comparable performance (AUROC 0.814 ± 0.10 and 0.814 ± 0.08, respectively). Across most classifiers, RF-based and RFE-LR feature selection did not improve diagnostic performance compared with N.S. Four radiomic features were identified by both RF and RFE-LR, being selected in at least two outer cross-validation folds.

Radiomics combined with machine learning demonstrated good performance for MI detection on noncontrast PMCT, emphasizing its potential as an objective, noninvasive tool for cause-of-death assessment, while suggesting a possible translational relevance for clinical contrast-free cardiac imaging.

Question Can MI detection on PMCT enable cause-of-death determination in forensic medicine and (clinical) pathology, particularly in settings without autopsy? Findings Radiomics-based machine learning on noncontrast PMCT showed good diagnostic performance for MI detection (mean AUROC 0.814-0.817). Relevance Statement Radiomics-based analysis of noncontrast PMCT shows promising potential for the detection of MI without the need for contrast agents or invasive procedures and may support forensic and pathological investigations by enhancing the diagnostic value of routinely acquired PMCT data. However, potential clinical applications require further validation.